linux/mm/vmscan.c
<<
>>
Prefs
   1/*
   2 *  linux/mm/vmscan.c
   3 *
   4 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5 *
   6 *  Swap reorganised 29.12.95, Stephen Tweedie.
   7 *  kswapd added: 7.1.96  sct
   8 *  Removed kswapd_ctl limits, and swap out as many pages as needed
   9 *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
  10 *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
  11 *  Multiqueue VM started 5.8.00, Rik van Riel.
  12 */
  13
  14#include <linux/mm.h>
  15#include <linux/module.h>
  16#include <linux/gfp.h>
  17#include <linux/kernel_stat.h>
  18#include <linux/swap.h>
  19#include <linux/pagemap.h>
  20#include <linux/init.h>
  21#include <linux/highmem.h>
  22#include <linux/vmpressure.h>
  23#include <linux/vmstat.h>
  24#include <linux/file.h>
  25#include <linux/writeback.h>
  26#include <linux/blkdev.h>
  27#include <linux/buffer_head.h>  /* for try_to_release_page(),
  28                                        buffer_heads_over_limit */
  29#include <linux/mm_inline.h>
  30#include <linux/backing-dev.h>
  31#include <linux/rmap.h>
  32#include <linux/topology.h>
  33#include <linux/cpu.h>
  34#include <linux/cpuset.h>
  35#include <linux/compaction.h>
  36#include <linux/notifier.h>
  37#include <linux/rwsem.h>
  38#include <linux/delay.h>
  39#include <linux/kthread.h>
  40#include <linux/freezer.h>
  41#include <linux/memcontrol.h>
  42#include <linux/delayacct.h>
  43#include <linux/sysctl.h>
  44#include <linux/oom.h>
  45#include <linux/prefetch.h>
  46
  47#include <asm/tlbflush.h>
  48#include <asm/div64.h>
  49
  50#include <linux/swapops.h>
  51
  52#include "internal.h"
  53
  54#define CREATE_TRACE_POINTS
  55#include <trace/events/vmscan.h>
  56
  57struct scan_control {
  58        /* Incremented by the number of inactive pages that were scanned */
  59        unsigned long nr_scanned;
  60
  61        /* Number of pages freed so far during a call to shrink_zones() */
  62        unsigned long nr_reclaimed;
  63
  64        /* How many pages shrink_list() should reclaim */
  65        unsigned long nr_to_reclaim;
  66
  67        unsigned long hibernation_mode;
  68
  69        /* This context's GFP mask */
  70        gfp_t gfp_mask;
  71
  72        int may_writepage;
  73
  74        /* Can mapped pages be reclaimed? */
  75        int may_unmap;
  76
  77        /* Can pages be swapped as part of reclaim? */
  78        int may_swap;
  79
  80        int order;
  81
  82        /* Scan (total_size >> priority) pages at once */
  83        int priority;
  84
  85        /*
  86         * The memory cgroup that hit its limit and as a result is the
  87         * primary target of this reclaim invocation.
  88         */
  89        struct mem_cgroup *target_mem_cgroup;
  90
  91        /*
  92         * Nodemask of nodes allowed by the caller. If NULL, all nodes
  93         * are scanned.
  94         */
  95        nodemask_t      *nodemask;
  96};
  97
  98#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  99
 100#ifdef ARCH_HAS_PREFETCH
 101#define prefetch_prev_lru_page(_page, _base, _field)                    \
 102        do {                                                            \
 103                if ((_page)->lru.prev != _base) {                       \
 104                        struct page *prev;                              \
 105                                                                        \
 106                        prev = lru_to_page(&(_page->lru));              \
 107                        prefetch(&prev->_field);                        \
 108                }                                                       \
 109        } while (0)
 110#else
 111#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 112#endif
 113
 114#ifdef ARCH_HAS_PREFETCHW
 115#define prefetchw_prev_lru_page(_page, _base, _field)                   \
 116        do {                                                            \
 117                if ((_page)->lru.prev != _base) {                       \
 118                        struct page *prev;                              \
 119                                                                        \
 120                        prev = lru_to_page(&(_page->lru));              \
 121                        prefetchw(&prev->_field);                       \
 122                }                                                       \
 123        } while (0)
 124#else
 125#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 126#endif
 127
 128/*
 129 * From 0 .. 100.  Higher means more swappy.
 130 */
 131int vm_swappiness = 60;
 132unsigned long vm_total_pages;   /* The total number of pages which the VM controls */
 133
 134static LIST_HEAD(shrinker_list);
 135static DECLARE_RWSEM(shrinker_rwsem);
 136
 137#ifdef CONFIG_MEMCG
 138static bool global_reclaim(struct scan_control *sc)
 139{
 140        return !sc->target_mem_cgroup;
 141}
 142#else
 143static bool global_reclaim(struct scan_control *sc)
 144{
 145        return true;
 146}
 147#endif
 148
 149static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
 150{
 151        if (!mem_cgroup_disabled())
 152                return mem_cgroup_get_lru_size(lruvec, lru);
 153
 154        return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
 155}
 156
 157/*
 158 * Add a shrinker callback to be called from the vm
 159 */
 160void register_shrinker(struct shrinker *shrinker)
 161{
 162        atomic_long_set(&shrinker->nr_in_batch, 0);
 163        down_write(&shrinker_rwsem);
 164        list_add_tail(&shrinker->list, &shrinker_list);
 165        up_write(&shrinker_rwsem);
 166}
 167EXPORT_SYMBOL(register_shrinker);
 168
 169/*
 170 * Remove one
 171 */
 172void unregister_shrinker(struct shrinker *shrinker)
 173{
 174        down_write(&shrinker_rwsem);
 175        list_del(&shrinker->list);
 176        up_write(&shrinker_rwsem);
 177}
 178EXPORT_SYMBOL(unregister_shrinker);
 179
 180static inline int do_shrinker_shrink(struct shrinker *shrinker,
 181                                     struct shrink_control *sc,
 182                                     unsigned long nr_to_scan)
 183{
 184        sc->nr_to_scan = nr_to_scan;
 185        return (*shrinker->shrink)(shrinker, sc);
 186}
 187
 188#define SHRINK_BATCH 128
 189/*
 190 * Call the shrink functions to age shrinkable caches
 191 *
 192 * Here we assume it costs one seek to replace a lru page and that it also
 193 * takes a seek to recreate a cache object.  With this in mind we age equal
 194 * percentages of the lru and ageable caches.  This should balance the seeks
 195 * generated by these structures.
 196 *
 197 * If the vm encountered mapped pages on the LRU it increase the pressure on
 198 * slab to avoid swapping.
 199 *
 200 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 201 *
 202 * `lru_pages' represents the number of on-LRU pages in all the zones which
 203 * are eligible for the caller's allocation attempt.  It is used for balancing
 204 * slab reclaim versus page reclaim.
 205 *
 206 * Returns the number of slab objects which we shrunk.
 207 */
 208unsigned long shrink_slab(struct shrink_control *shrink,
 209                          unsigned long nr_pages_scanned,
 210                          unsigned long lru_pages)
 211{
 212        struct shrinker *shrinker;
 213        unsigned long ret = 0;
 214
 215        if (nr_pages_scanned == 0)
 216                nr_pages_scanned = SWAP_CLUSTER_MAX;
 217
 218        if (!down_read_trylock(&shrinker_rwsem)) {
 219                /* Assume we'll be able to shrink next time */
 220                ret = 1;
 221                goto out;
 222        }
 223
 224        list_for_each_entry(shrinker, &shrinker_list, list) {
 225                unsigned long long delta;
 226                long total_scan;
 227                long max_pass;
 228                int shrink_ret = 0;
 229                long nr;
 230                long new_nr;
 231                long batch_size = shrinker->batch ? shrinker->batch
 232                                                  : SHRINK_BATCH;
 233
 234                max_pass = do_shrinker_shrink(shrinker, shrink, 0);
 235                if (max_pass <= 0)
 236                        continue;
 237
 238                /*
 239                 * copy the current shrinker scan count into a local variable
 240                 * and zero it so that other concurrent shrinker invocations
 241                 * don't also do this scanning work.
 242                 */
 243                nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
 244
 245                total_scan = nr;
 246                delta = (4 * nr_pages_scanned) / shrinker->seeks;
 247                delta *= max_pass;
 248                do_div(delta, lru_pages + 1);
 249                total_scan += delta;
 250                if (total_scan < 0) {
 251                        printk(KERN_ERR "shrink_slab: %pF negative objects to "
 252                               "delete nr=%ld\n",
 253                               shrinker->shrink, total_scan);
 254                        total_scan = max_pass;
 255                }
 256
 257                /*
 258                 * We need to avoid excessive windup on filesystem shrinkers
 259                 * due to large numbers of GFP_NOFS allocations causing the
 260                 * shrinkers to return -1 all the time. This results in a large
 261                 * nr being built up so when a shrink that can do some work
 262                 * comes along it empties the entire cache due to nr >>>
 263                 * max_pass.  This is bad for sustaining a working set in
 264                 * memory.
 265                 *
 266                 * Hence only allow the shrinker to scan the entire cache when
 267                 * a large delta change is calculated directly.
 268                 */
 269                if (delta < max_pass / 4)
 270                        total_scan = min(total_scan, max_pass / 2);
 271
 272                /*
 273                 * Avoid risking looping forever due to too large nr value:
 274                 * never try to free more than twice the estimate number of
 275                 * freeable entries.
 276                 */
 277                if (total_scan > max_pass * 2)
 278                        total_scan = max_pass * 2;
 279
 280                trace_mm_shrink_slab_start(shrinker, shrink, nr,
 281                                        nr_pages_scanned, lru_pages,
 282                                        max_pass, delta, total_scan);
 283
 284                while (total_scan >= batch_size) {
 285                        int nr_before;
 286
 287                        nr_before = do_shrinker_shrink(shrinker, shrink, 0);
 288                        shrink_ret = do_shrinker_shrink(shrinker, shrink,
 289                                                        batch_size);
 290                        if (shrink_ret == -1)
 291                                break;
 292                        if (shrink_ret < nr_before)
 293                                ret += nr_before - shrink_ret;
 294                        count_vm_events(SLABS_SCANNED, batch_size);
 295                        total_scan -= batch_size;
 296
 297                        cond_resched();
 298                }
 299
 300                /*
 301                 * move the unused scan count back into the shrinker in a
 302                 * manner that handles concurrent updates. If we exhausted the
 303                 * scan, there is no need to do an update.
 304                 */
 305                if (total_scan > 0)
 306                        new_nr = atomic_long_add_return(total_scan,
 307                                        &shrinker->nr_in_batch);
 308                else
 309                        new_nr = atomic_long_read(&shrinker->nr_in_batch);
 310
 311                trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
 312        }
 313        up_read(&shrinker_rwsem);
 314out:
 315        cond_resched();
 316        return ret;
 317}
 318
 319static inline int is_page_cache_freeable(struct page *page)
 320{
 321        /*
 322         * A freeable page cache page is referenced only by the caller
 323         * that isolated the page, the page cache radix tree and
 324         * optional buffer heads at page->private.
 325         */
 326        return page_count(page) - page_has_private(page) == 2;
 327}
 328
 329static int may_write_to_queue(struct backing_dev_info *bdi,
 330                              struct scan_control *sc)
 331{
 332        if (current->flags & PF_SWAPWRITE)
 333                return 1;
 334        if (!bdi_write_congested(bdi))
 335                return 1;
 336        if (bdi == current->backing_dev_info)
 337                return 1;
 338        return 0;
 339}
 340
 341/*
 342 * We detected a synchronous write error writing a page out.  Probably
 343 * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 344 * fsync(), msync() or close().
 345 *
 346 * The tricky part is that after writepage we cannot touch the mapping: nothing
 347 * prevents it from being freed up.  But we have a ref on the page and once
 348 * that page is locked, the mapping is pinned.
 349 *
 350 * We're allowed to run sleeping lock_page() here because we know the caller has
 351 * __GFP_FS.
 352 */
 353static void handle_write_error(struct address_space *mapping,
 354                                struct page *page, int error)
 355{
 356        lock_page(page);
 357        if (page_mapping(page) == mapping)
 358                mapping_set_error(mapping, error);
 359        unlock_page(page);
 360}
 361
 362/* possible outcome of pageout() */
 363typedef enum {
 364        /* failed to write page out, page is locked */
 365        PAGE_KEEP,
 366        /* move page to the active list, page is locked */
 367        PAGE_ACTIVATE,
 368        /* page has been sent to the disk successfully, page is unlocked */
 369        PAGE_SUCCESS,
 370        /* page is clean and locked */
 371        PAGE_CLEAN,
 372} pageout_t;
 373
 374/*
 375 * pageout is called by shrink_page_list() for each dirty page.
 376 * Calls ->writepage().
 377 */
 378static pageout_t pageout(struct page *page, struct address_space *mapping,
 379                         struct scan_control *sc)
 380{
 381        /*
 382         * If the page is dirty, only perform writeback if that write
 383         * will be non-blocking.  To prevent this allocation from being
 384         * stalled by pagecache activity.  But note that there may be
 385         * stalls if we need to run get_block().  We could test
 386         * PagePrivate for that.
 387         *
 388         * If this process is currently in __generic_file_aio_write() against
 389         * this page's queue, we can perform writeback even if that
 390         * will block.
 391         *
 392         * If the page is swapcache, write it back even if that would
 393         * block, for some throttling. This happens by accident, because
 394         * swap_backing_dev_info is bust: it doesn't reflect the
 395         * congestion state of the swapdevs.  Easy to fix, if needed.
 396         */
 397        if (!is_page_cache_freeable(page))
 398                return PAGE_KEEP;
 399        if (!mapping) {
 400                /*
 401                 * Some data journaling orphaned pages can have
 402                 * page->mapping == NULL while being dirty with clean buffers.
 403                 */
 404                if (page_has_private(page)) {
 405                        if (try_to_free_buffers(page)) {
 406                                ClearPageDirty(page);
 407                                printk("%s: orphaned page\n", __func__);
 408                                return PAGE_CLEAN;
 409                        }
 410                }
 411                return PAGE_KEEP;
 412        }
 413        if (mapping->a_ops->writepage == NULL)
 414                return PAGE_ACTIVATE;
 415        if (!may_write_to_queue(mapping->backing_dev_info, sc))
 416                return PAGE_KEEP;
 417
 418        if (clear_page_dirty_for_io(page)) {
 419                int res;
 420                struct writeback_control wbc = {
 421                        .sync_mode = WB_SYNC_NONE,
 422                        .nr_to_write = SWAP_CLUSTER_MAX,
 423                        .range_start = 0,
 424                        .range_end = LLONG_MAX,
 425                        .for_reclaim = 1,
 426                };
 427
 428                SetPageReclaim(page);
 429                res = mapping->a_ops->writepage(page, &wbc);
 430                if (res < 0)
 431                        handle_write_error(mapping, page, res);
 432                if (res == AOP_WRITEPAGE_ACTIVATE) {
 433                        ClearPageReclaim(page);
 434                        return PAGE_ACTIVATE;
 435                }
 436
 437                if (!PageWriteback(page)) {
 438                        /* synchronous write or broken a_ops? */
 439                        ClearPageReclaim(page);
 440                }
 441                trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
 442                inc_zone_page_state(page, NR_VMSCAN_WRITE);
 443                return PAGE_SUCCESS;
 444        }
 445
 446        return PAGE_CLEAN;
 447}
 448
 449/*
 450 * Same as remove_mapping, but if the page is removed from the mapping, it
 451 * gets returned with a refcount of 0.
 452 */
 453static int __remove_mapping(struct address_space *mapping, struct page *page)
 454{
 455        BUG_ON(!PageLocked(page));
 456        BUG_ON(mapping != page_mapping(page));
 457
 458        spin_lock_irq(&mapping->tree_lock);
 459        /*
 460         * The non racy check for a busy page.
 461         *
 462         * Must be careful with the order of the tests. When someone has
 463         * a ref to the page, it may be possible that they dirty it then
 464         * drop the reference. So if PageDirty is tested before page_count
 465         * here, then the following race may occur:
 466         *
 467         * get_user_pages(&page);
 468         * [user mapping goes away]
 469         * write_to(page);
 470         *                              !PageDirty(page)    [good]
 471         * SetPageDirty(page);
 472         * put_page(page);
 473         *                              !page_count(page)   [good, discard it]
 474         *
 475         * [oops, our write_to data is lost]
 476         *
 477         * Reversing the order of the tests ensures such a situation cannot
 478         * escape unnoticed. The smp_rmb is needed to ensure the page->flags
 479         * load is not satisfied before that of page->_count.
 480         *
 481         * Note that if SetPageDirty is always performed via set_page_dirty,
 482         * and thus under tree_lock, then this ordering is not required.
 483         */
 484        if (!page_freeze_refs(page, 2))
 485                goto cannot_free;
 486        /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
 487        if (unlikely(PageDirty(page))) {
 488                page_unfreeze_refs(page, 2);
 489                goto cannot_free;
 490        }
 491
 492        if (PageSwapCache(page)) {
 493                swp_entry_t swap = { .val = page_private(page) };
 494                __delete_from_swap_cache(page);
 495                spin_unlock_irq(&mapping->tree_lock);
 496                swapcache_free(swap, page);
 497        } else {
 498                void (*freepage)(struct page *);
 499
 500                freepage = mapping->a_ops->freepage;
 501
 502                __delete_from_page_cache(page);
 503                spin_unlock_irq(&mapping->tree_lock);
 504                mem_cgroup_uncharge_cache_page(page);
 505
 506                if (freepage != NULL)
 507                        freepage(page);
 508        }
 509
 510        return 1;
 511
 512cannot_free:
 513        spin_unlock_irq(&mapping->tree_lock);
 514        return 0;
 515}
 516
 517/*
 518 * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
 519 * someone else has a ref on the page, abort and return 0.  If it was
 520 * successfully detached, return 1.  Assumes the caller has a single ref on
 521 * this page.
 522 */
 523int remove_mapping(struct address_space *mapping, struct page *page)
 524{
 525        if (__remove_mapping(mapping, page)) {
 526                /*
 527                 * Unfreezing the refcount with 1 rather than 2 effectively
 528                 * drops the pagecache ref for us without requiring another
 529                 * atomic operation.
 530                 */
 531                page_unfreeze_refs(page, 1);
 532                return 1;
 533        }
 534        return 0;
 535}
 536
 537/**
 538 * putback_lru_page - put previously isolated page onto appropriate LRU list
 539 * @page: page to be put back to appropriate lru list
 540 *
 541 * Add previously isolated @page to appropriate LRU list.
 542 * Page may still be unevictable for other reasons.
 543 *
 544 * lru_lock must not be held, interrupts must be enabled.
 545 */
 546void putback_lru_page(struct page *page)
 547{
 548        int lru;
 549        int active = !!TestClearPageActive(page);
 550        int was_unevictable = PageUnevictable(page);
 551
 552        VM_BUG_ON(PageLRU(page));
 553
 554redo:
 555        ClearPageUnevictable(page);
 556
 557        if (page_evictable(page)) {
 558                /*
 559                 * For evictable pages, we can use the cache.
 560                 * In event of a race, worst case is we end up with an
 561                 * unevictable page on [in]active list.
 562                 * We know how to handle that.
 563                 */
 564                lru = active + page_lru_base_type(page);
 565                lru_cache_add_lru(page, lru);
 566        } else {
 567                /*
 568                 * Put unevictable pages directly on zone's unevictable
 569                 * list.
 570                 */
 571                lru = LRU_UNEVICTABLE;
 572                add_page_to_unevictable_list(page);
 573                /*
 574                 * When racing with an mlock or AS_UNEVICTABLE clearing
 575                 * (page is unlocked) make sure that if the other thread
 576                 * does not observe our setting of PG_lru and fails
 577                 * isolation/check_move_unevictable_pages,
 578                 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
 579                 * the page back to the evictable list.
 580                 *
 581                 * The other side is TestClearPageMlocked() or shmem_lock().
 582                 */
 583                smp_mb();
 584        }
 585
 586        /*
 587         * page's status can change while we move it among lru. If an evictable
 588         * page is on unevictable list, it never be freed. To avoid that,
 589         * check after we added it to the list, again.
 590         */
 591        if (lru == LRU_UNEVICTABLE && page_evictable(page)) {
 592                if (!isolate_lru_page(page)) {
 593                        put_page(page);
 594                        goto redo;
 595                }
 596                /* This means someone else dropped this page from LRU
 597                 * So, it will be freed or putback to LRU again. There is
 598                 * nothing to do here.
 599                 */
 600        }
 601
 602        if (was_unevictable && lru != LRU_UNEVICTABLE)
 603                count_vm_event(UNEVICTABLE_PGRESCUED);
 604        else if (!was_unevictable && lru == LRU_UNEVICTABLE)
 605                count_vm_event(UNEVICTABLE_PGCULLED);
 606
 607        put_page(page);         /* drop ref from isolate */
 608}
 609
 610enum page_references {
 611        PAGEREF_RECLAIM,
 612        PAGEREF_RECLAIM_CLEAN,
 613        PAGEREF_KEEP,
 614        PAGEREF_ACTIVATE,
 615};
 616
 617static enum page_references page_check_references(struct page *page,
 618                                                  struct scan_control *sc)
 619{
 620        int referenced_ptes, referenced_page;
 621        unsigned long vm_flags;
 622
 623        referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
 624                                          &vm_flags);
 625        referenced_page = TestClearPageReferenced(page);
 626
 627        /*
 628         * Mlock lost the isolation race with us.  Let try_to_unmap()
 629         * move the page to the unevictable list.
 630         */
 631        if (vm_flags & VM_LOCKED)
 632                return PAGEREF_RECLAIM;
 633
 634        if (referenced_ptes) {
 635                if (PageSwapBacked(page))
 636                        return PAGEREF_ACTIVATE;
 637                /*
 638                 * All mapped pages start out with page table
 639                 * references from the instantiating fault, so we need
 640                 * to look twice if a mapped file page is used more
 641                 * than once.
 642                 *
 643                 * Mark it and spare it for another trip around the
 644                 * inactive list.  Another page table reference will
 645                 * lead to its activation.
 646                 *
 647                 * Note: the mark is set for activated pages as well
 648                 * so that recently deactivated but used pages are
 649                 * quickly recovered.
 650                 */
 651                SetPageReferenced(page);
 652
 653                if (referenced_page || referenced_ptes > 1)
 654                        return PAGEREF_ACTIVATE;
 655
 656                /*
 657                 * Activate file-backed executable pages after first usage.
 658                 */
 659                if (vm_flags & VM_EXEC)
 660                        return PAGEREF_ACTIVATE;
 661
 662                return PAGEREF_KEEP;
 663        }
 664
 665        /* Reclaim if clean, defer dirty pages to writeback */
 666        if (referenced_page && !PageSwapBacked(page))
 667                return PAGEREF_RECLAIM_CLEAN;
 668
 669        return PAGEREF_RECLAIM;
 670}
 671
 672/*
 673 * shrink_page_list() returns the number of reclaimed pages
 674 */
 675static unsigned long shrink_page_list(struct list_head *page_list,
 676                                      struct zone *zone,
 677                                      struct scan_control *sc,
 678                                      enum ttu_flags ttu_flags,
 679                                      unsigned long *ret_nr_dirty,
 680                                      unsigned long *ret_nr_writeback,
 681                                      bool force_reclaim)
 682{
 683        LIST_HEAD(ret_pages);
 684        LIST_HEAD(free_pages);
 685        int pgactivate = 0;
 686        unsigned long nr_dirty = 0;
 687        unsigned long nr_congested = 0;
 688        unsigned long nr_reclaimed = 0;
 689        unsigned long nr_writeback = 0;
 690
 691        cond_resched();
 692
 693        mem_cgroup_uncharge_start();
 694        while (!list_empty(page_list)) {
 695                struct address_space *mapping;
 696                struct page *page;
 697                int may_enter_fs;
 698                enum page_references references = PAGEREF_RECLAIM_CLEAN;
 699
 700                cond_resched();
 701
 702                page = lru_to_page(page_list);
 703                list_del(&page->lru);
 704
 705                if (!trylock_page(page))
 706                        goto keep;
 707
 708                VM_BUG_ON(PageActive(page));
 709                VM_BUG_ON(page_zone(page) != zone);
 710
 711                sc->nr_scanned++;
 712
 713                if (unlikely(!page_evictable(page)))
 714                        goto cull_mlocked;
 715
 716                if (!sc->may_unmap && page_mapped(page))
 717                        goto keep_locked;
 718
 719                /* Double the slab pressure for mapped and swapcache pages */
 720                if (page_mapped(page) || PageSwapCache(page))
 721                        sc->nr_scanned++;
 722
 723                may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 724                        (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 725
 726                if (PageWriteback(page)) {
 727                        /*
 728                         * memcg doesn't have any dirty pages throttling so we
 729                         * could easily OOM just because too many pages are in
 730                         * writeback and there is nothing else to reclaim.
 731                         *
 732                         * Check __GFP_IO, certainly because a loop driver
 733                         * thread might enter reclaim, and deadlock if it waits
 734                         * on a page for which it is needed to do the write
 735                         * (loop masks off __GFP_IO|__GFP_FS for this reason);
 736                         * but more thought would probably show more reasons.
 737                         *
 738                         * Don't require __GFP_FS, since we're not going into
 739                         * the FS, just waiting on its writeback completion.
 740                         * Worryingly, ext4 gfs2 and xfs allocate pages with
 741                         * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so
 742                         * testing may_enter_fs here is liable to OOM on them.
 743                         */
 744                        if (global_reclaim(sc) ||
 745                            !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
 746                                /*
 747                                 * This is slightly racy - end_page_writeback()
 748                                 * might have just cleared PageReclaim, then
 749                                 * setting PageReclaim here end up interpreted
 750                                 * as PageReadahead - but that does not matter
 751                                 * enough to care.  What we do want is for this
 752                                 * page to have PageReclaim set next time memcg
 753                                 * reclaim reaches the tests above, so it will
 754                                 * then wait_on_page_writeback() to avoid OOM;
 755                                 * and it's also appropriate in global reclaim.
 756                                 */
 757                                SetPageReclaim(page);
 758                                nr_writeback++;
 759                                goto keep_locked;
 760                        }
 761                        wait_on_page_writeback(page);
 762                }
 763
 764                if (!force_reclaim)
 765                        references = page_check_references(page, sc);
 766
 767                switch (references) {
 768                case PAGEREF_ACTIVATE:
 769                        goto activate_locked;
 770                case PAGEREF_KEEP:
 771                        goto keep_locked;
 772                case PAGEREF_RECLAIM:
 773                case PAGEREF_RECLAIM_CLEAN:
 774                        ; /* try to reclaim the page below */
 775                }
 776
 777                /*
 778                 * Anonymous process memory has backing store?
 779                 * Try to allocate it some swap space here.
 780                 */
 781                if (PageAnon(page) && !PageSwapCache(page)) {
 782                        if (!(sc->gfp_mask & __GFP_IO))
 783                                goto keep_locked;
 784                        if (!add_to_swap(page, page_list))
 785                                goto activate_locked;
 786                        may_enter_fs = 1;
 787                }
 788
 789                mapping = page_mapping(page);
 790
 791                /*
 792                 * The page is mapped into the page tables of one or more
 793                 * processes. Try to unmap it here.
 794                 */
 795                if (page_mapped(page) && mapping) {
 796                        switch (try_to_unmap(page, ttu_flags)) {
 797                        case SWAP_FAIL:
 798                                goto activate_locked;
 799                        case SWAP_AGAIN:
 800                                goto keep_locked;
 801                        case SWAP_MLOCK:
 802                                goto cull_mlocked;
 803                        case SWAP_SUCCESS:
 804                                ; /* try to free the page below */
 805                        }
 806                }
 807
 808                if (PageDirty(page)) {
 809                        nr_dirty++;
 810
 811                        /*
 812                         * Only kswapd can writeback filesystem pages to
 813                         * avoid risk of stack overflow but do not writeback
 814                         * unless under significant pressure.
 815                         */
 816                        if (page_is_file_cache(page) &&
 817                                        (!current_is_kswapd() ||
 818                                         sc->priority >= DEF_PRIORITY - 2)) {
 819                                /*
 820                                 * Immediately reclaim when written back.
 821                                 * Similar in principal to deactivate_page()
 822                                 * except we already have the page isolated
 823                                 * and know it's dirty
 824                                 */
 825                                inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
 826                                SetPageReclaim(page);
 827
 828                                goto keep_locked;
 829                        }
 830
 831                        if (references == PAGEREF_RECLAIM_CLEAN)
 832                                goto keep_locked;
 833                        if (!may_enter_fs)
 834                                goto keep_locked;
 835                        if (!sc->may_writepage)
 836                                goto keep_locked;
 837
 838                        /* Page is dirty, try to write it out here */
 839                        switch (pageout(page, mapping, sc)) {
 840                        case PAGE_KEEP:
 841                                nr_congested++;
 842                                goto keep_locked;
 843                        case PAGE_ACTIVATE:
 844                                goto activate_locked;
 845                        case PAGE_SUCCESS:
 846                                if (PageWriteback(page))
 847                                        goto keep;
 848                                if (PageDirty(page))
 849                                        goto keep;
 850
 851                                /*
 852                                 * A synchronous write - probably a ramdisk.  Go
 853                                 * ahead and try to reclaim the page.
 854                                 */
 855                                if (!trylock_page(page))
 856                                        goto keep;
 857                                if (PageDirty(page) || PageWriteback(page))
 858                                        goto keep_locked;
 859                                mapping = page_mapping(page);
 860                        case PAGE_CLEAN:
 861                                ; /* try to free the page below */
 862                        }
 863                }
 864
 865                /*
 866                 * If the page has buffers, try to free the buffer mappings
 867                 * associated with this page. If we succeed we try to free
 868                 * the page as well.
 869                 *
 870                 * We do this even if the page is PageDirty().
 871                 * try_to_release_page() does not perform I/O, but it is
 872                 * possible for a page to have PageDirty set, but it is actually
 873                 * clean (all its buffers are clean).  This happens if the
 874                 * buffers were written out directly, with submit_bh(). ext3
 875                 * will do this, as well as the blockdev mapping.
 876                 * try_to_release_page() will discover that cleanness and will
 877                 * drop the buffers and mark the page clean - it can be freed.
 878                 *
 879                 * Rarely, pages can have buffers and no ->mapping.  These are
 880                 * the pages which were not successfully invalidated in
 881                 * truncate_complete_page().  We try to drop those buffers here
 882                 * and if that worked, and the page is no longer mapped into
 883                 * process address space (page_count == 1) it can be freed.
 884                 * Otherwise, leave the page on the LRU so it is swappable.
 885                 */
 886                if (page_has_private(page)) {
 887                        if (!try_to_release_page(page, sc->gfp_mask))
 888                                goto activate_locked;
 889                        if (!mapping && page_count(page) == 1) {
 890                                unlock_page(page);
 891                                if (put_page_testzero(page))
 892                                        goto free_it;
 893                                else {
 894                                        /*
 895                                         * rare race with speculative reference.
 896                                         * the speculative reference will free
 897                                         * this page shortly, so we may
 898                                         * increment nr_reclaimed here (and
 899                                         * leave it off the LRU).
 900                                         */
 901                                        nr_reclaimed++;
 902                                        continue;
 903                                }
 904                        }
 905                }
 906
 907                if (!mapping || !__remove_mapping(mapping, page))
 908                        goto keep_locked;
 909
 910                /*
 911                 * At this point, we have no other references and there is
 912                 * no way to pick any more up (removed from LRU, removed
 913                 * from pagecache). Can use non-atomic bitops now (and
 914                 * we obviously don't have to worry about waking up a process
 915                 * waiting on the page lock, because there are no references.
 916                 */
 917                __clear_page_locked(page);
 918free_it:
 919                nr_reclaimed++;
 920
 921                /*
 922                 * Is there need to periodically free_page_list? It would
 923                 * appear not as the counts should be low
 924                 */
 925                list_add(&page->lru, &free_pages);
 926                continue;
 927
 928cull_mlocked:
 929                if (PageSwapCache(page))
 930                        try_to_free_swap(page);
 931                unlock_page(page);
 932                putback_lru_page(page);
 933                continue;
 934
 935activate_locked:
 936                /* Not a candidate for swapping, so reclaim swap space. */
 937                if (PageSwapCache(page) && vm_swap_full())
 938                        try_to_free_swap(page);
 939                VM_BUG_ON(PageActive(page));
 940                SetPageActive(page);
 941                pgactivate++;
 942keep_locked:
 943                unlock_page(page);
 944keep:
 945                list_add(&page->lru, &ret_pages);
 946                VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
 947        }
 948
 949        /*
 950         * Tag a zone as congested if all the dirty pages encountered were
 951         * backed by a congested BDI. In this case, reclaimers should just
 952         * back off and wait for congestion to clear because further reclaim
 953         * will encounter the same problem
 954         */
 955        if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
 956                zone_set_flag(zone, ZONE_CONGESTED);
 957
 958        free_hot_cold_page_list(&free_pages, 1);
 959
 960        list_splice(&ret_pages, page_list);
 961        count_vm_events(PGACTIVATE, pgactivate);
 962        mem_cgroup_uncharge_end();
 963        *ret_nr_dirty += nr_dirty;
 964        *ret_nr_writeback += nr_writeback;
 965        return nr_reclaimed;
 966}
 967
 968unsigned long reclaim_clean_pages_from_list(struct zone *zone,
 969                                            struct list_head *page_list)
 970{
 971        struct scan_control sc = {
 972                .gfp_mask = GFP_KERNEL,
 973                .priority = DEF_PRIORITY,
 974                .may_unmap = 1,
 975        };
 976        unsigned long ret, dummy1, dummy2;
 977        struct page *page, *next;
 978        LIST_HEAD(clean_pages);
 979
 980        list_for_each_entry_safe(page, next, page_list, lru) {
 981                if (page_is_file_cache(page) && !PageDirty(page)) {
 982                        ClearPageActive(page);
 983                        list_move(&page->lru, &clean_pages);
 984                }
 985        }
 986
 987        ret = shrink_page_list(&clean_pages, zone, &sc,
 988                                TTU_UNMAP|TTU_IGNORE_ACCESS,
 989                                &dummy1, &dummy2, true);
 990        list_splice(&clean_pages, page_list);
 991        __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
 992        return ret;
 993}
 994
 995/*
 996 * Attempt to remove the specified page from its LRU.  Only take this page
 997 * if it is of the appropriate PageActive status.  Pages which are being
 998 * freed elsewhere are also ignored.
 999 *
1000 * page:        page to consider
1001 * mode:        one of the LRU isolation modes defined above
1002 *
1003 * returns 0 on success, -ve errno on failure.
1004 */
1005int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1006{
1007        int ret = -EINVAL;
1008
1009        /* Only take pages on the LRU. */
1010        if (!PageLRU(page))
1011                return ret;
1012
1013        /* Compaction should not handle unevictable pages but CMA can do so */
1014        if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1015                return ret;
1016
1017        ret = -EBUSY;
1018
1019        /*
1020         * To minimise LRU disruption, the caller can indicate that it only
1021         * wants to isolate pages it will be able to operate on without
1022         * blocking - clean pages for the most part.
1023         *
1024         * ISOLATE_CLEAN means that only clean pages should be isolated. This
1025         * is used by reclaim when it is cannot write to backing storage
1026         *
1027         * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1028         * that it is possible to migrate without blocking
1029         */
1030        if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1031                /* All the caller can do on PageWriteback is block */
1032                if (PageWriteback(page))
1033                        return ret;
1034
1035                if (PageDirty(page)) {
1036                        struct address_space *mapping;
1037
1038                        /* ISOLATE_CLEAN means only clean pages */
1039                        if (mode & ISOLATE_CLEAN)
1040                                return ret;
1041
1042                        /*
1043                         * Only pages without mappings or that have a
1044                         * ->migratepage callback are possible to migrate
1045                         * without blocking
1046                         */
1047                        mapping = page_mapping(page);
1048                        if (mapping && !mapping->a_ops->migratepage)
1049                                return ret;
1050                }
1051        }
1052
1053        if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1054                return ret;
1055
1056        if (likely(get_page_unless_zero(page))) {
1057                /*
1058                 * Be careful not to clear PageLRU until after we're
1059                 * sure the page is not being freed elsewhere -- the
1060                 * page release code relies on it.
1061                 */
1062                ClearPageLRU(page);
1063                ret = 0;
1064        }
1065
1066        return ret;
1067}
1068
1069/*
1070 * zone->lru_lock is heavily contended.  Some of the functions that
1071 * shrink the lists perform better by taking out a batch of pages
1072 * and working on them outside the LRU lock.
1073 *
1074 * For pagecache intensive workloads, this function is the hottest
1075 * spot in the kernel (apart from copy_*_user functions).
1076 *
1077 * Appropriate locks must be held before calling this function.
1078 *
1079 * @nr_to_scan: The number of pages to look through on the list.
1080 * @lruvec:     The LRU vector to pull pages from.
1081 * @dst:        The temp list to put pages on to.
1082 * @nr_scanned: The number of pages that were scanned.
1083 * @sc:         The scan_control struct for this reclaim session
1084 * @mode:       One of the LRU isolation modes
1085 * @lru:        LRU list id for isolating
1086 *
1087 * returns how many pages were moved onto *@dst.
1088 */
1089static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1090                struct lruvec *lruvec, struct list_head *dst,
1091                unsigned long *nr_scanned, struct scan_control *sc,
1092                isolate_mode_t mode, enum lru_list lru)
1093{
1094        struct list_head *src = &lruvec->lists[lru];
1095        unsigned long nr_taken = 0;
1096        unsigned long scan;
1097
1098        for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1099                struct page *page;
1100                int nr_pages;
1101
1102                page = lru_to_page(src);
1103                prefetchw_prev_lru_page(page, src, flags);
1104
1105                VM_BUG_ON(!PageLRU(page));
1106
1107                switch (__isolate_lru_page(page, mode)) {
1108                case 0:
1109                        nr_pages = hpage_nr_pages(page);
1110                        mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1111                        list_move(&page->lru, dst);
1112                        nr_taken += nr_pages;
1113                        break;
1114
1115                case -EBUSY:
1116                        /* else it is being freed elsewhere */
1117                        list_move(&page->lru, src);
1118                        continue;
1119
1120                default:
1121                        BUG();
1122                }
1123        }
1124
1125        *nr_scanned = scan;
1126        trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1127                                    nr_taken, mode, is_file_lru(lru));
1128        return nr_taken;
1129}
1130
1131/**
1132 * isolate_lru_page - tries to isolate a page from its LRU list
1133 * @page: page to isolate from its LRU list
1134 *
1135 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1136 * vmstat statistic corresponding to whatever LRU list the page was on.
1137 *
1138 * Returns 0 if the page was removed from an LRU list.
1139 * Returns -EBUSY if the page was not on an LRU list.
1140 *
1141 * The returned page will have PageLRU() cleared.  If it was found on
1142 * the active list, it will have PageActive set.  If it was found on
1143 * the unevictable list, it will have the PageUnevictable bit set. That flag
1144 * may need to be cleared by the caller before letting the page go.
1145 *
1146 * The vmstat statistic corresponding to the list on which the page was
1147 * found will be decremented.
1148 *
1149 * Restrictions:
1150 * (1) Must be called with an elevated refcount on the page. This is a
1151 *     fundamentnal difference from isolate_lru_pages (which is called
1152 *     without a stable reference).
1153 * (2) the lru_lock must not be held.
1154 * (3) interrupts must be enabled.
1155 */
1156int isolate_lru_page(struct page *page)
1157{
1158        int ret = -EBUSY;
1159
1160        VM_BUG_ON(!page_count(page));
1161
1162        if (PageLRU(page)) {
1163                struct zone *zone = page_zone(page);
1164                struct lruvec *lruvec;
1165
1166                spin_lock_irq(&zone->lru_lock);
1167                lruvec = mem_cgroup_page_lruvec(page, zone);
1168                if (PageLRU(page)) {
1169                        int lru = page_lru(page);
1170                        get_page(page);
1171                        ClearPageLRU(page);
1172                        del_page_from_lru_list(page, lruvec, lru);
1173                        ret = 0;
1174                }
1175                spin_unlock_irq(&zone->lru_lock);
1176        }
1177        return ret;
1178}
1179
1180/*
1181 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1182 * then get resheduled. When there are massive number of tasks doing page
1183 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1184 * the LRU list will go small and be scanned faster than necessary, leading to
1185 * unnecessary swapping, thrashing and OOM.
1186 */
1187static int too_many_isolated(struct zone *zone, int file,
1188                struct scan_control *sc)
1189{
1190        unsigned long inactive, isolated;
1191
1192        if (current_is_kswapd())
1193                return 0;
1194
1195        if (!global_reclaim(sc))
1196                return 0;
1197
1198        if (file) {
1199                inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1200                isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1201        } else {
1202                inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1203                isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1204        }
1205
1206        /*
1207         * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1208         * won't get blocked by normal direct-reclaimers, forming a circular
1209         * deadlock.
1210         */
1211        if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1212                inactive >>= 3;
1213
1214        return isolated > inactive;
1215}
1216
1217static noinline_for_stack void
1218putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1219{
1220        struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1221        struct zone *zone = lruvec_zone(lruvec);
1222        LIST_HEAD(pages_to_free);
1223
1224        /*
1225         * Put back any unfreeable pages.
1226         */
1227        while (!list_empty(page_list)) {
1228                struct page *page = lru_to_page(page_list);
1229                int lru;
1230
1231                VM_BUG_ON(PageLRU(page));
1232                list_del(&page->lru);
1233                if (unlikely(!page_evictable(page))) {
1234                        spin_unlock_irq(&zone->lru_lock);
1235                        putback_lru_page(page);
1236                        spin_lock_irq(&zone->lru_lock);
1237                        continue;
1238                }
1239
1240                lruvec = mem_cgroup_page_lruvec(page, zone);
1241
1242                SetPageLRU(page);
1243                lru = page_lru(page);
1244                add_page_to_lru_list(page, lruvec, lru);
1245
1246                if (is_active_lru(lru)) {
1247                        int file = is_file_lru(lru);
1248                        int numpages = hpage_nr_pages(page);
1249                        reclaim_stat->recent_rotated[file] += numpages;
1250                }
1251                if (put_page_testzero(page)) {
1252                        __ClearPageLRU(page);
1253                        __ClearPageActive(page);
1254                        del_page_from_lru_list(page, lruvec, lru);
1255
1256                        if (unlikely(PageCompound(page))) {
1257                                spin_unlock_irq(&zone->lru_lock);
1258                                (*get_compound_page_dtor(page))(page);
1259                                spin_lock_irq(&zone->lru_lock);
1260                        } else
1261                                list_add(&page->lru, &pages_to_free);
1262                }
1263        }
1264
1265        /*
1266         * To save our caller's stack, now use input list for pages to free.
1267         */
1268        list_splice(&pages_to_free, page_list);
1269}
1270
1271/*
1272 * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1273 * of reclaimed pages
1274 */
1275static noinline_for_stack unsigned long
1276shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1277                     struct scan_control *sc, enum lru_list lru)
1278{
1279        LIST_HEAD(page_list);
1280        unsigned long nr_scanned;
1281        unsigned long nr_reclaimed = 0;
1282        unsigned long nr_taken;
1283        unsigned long nr_dirty = 0;
1284        unsigned long nr_writeback = 0;
1285        isolate_mode_t isolate_mode = 0;
1286        int file = is_file_lru(lru);
1287        struct zone *zone = lruvec_zone(lruvec);
1288        struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1289
1290        while (unlikely(too_many_isolated(zone, file, sc))) {
1291                congestion_wait(BLK_RW_ASYNC, HZ/10);
1292
1293                /* We are about to die and free our memory. Return now. */
1294                if (fatal_signal_pending(current))
1295                        return SWAP_CLUSTER_MAX;
1296        }
1297
1298        lru_add_drain();
1299
1300        if (!sc->may_unmap)
1301                isolate_mode |= ISOLATE_UNMAPPED;
1302        if (!sc->may_writepage)
1303                isolate_mode |= ISOLATE_CLEAN;
1304
1305        spin_lock_irq(&zone->lru_lock);
1306
1307        nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1308                                     &nr_scanned, sc, isolate_mode, lru);
1309
1310        __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1311        __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1312
1313        if (global_reclaim(sc)) {
1314                zone->pages_scanned += nr_scanned;
1315                if (current_is_kswapd())
1316                        __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1317                else
1318                        __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1319        }
1320        spin_unlock_irq(&zone->lru_lock);
1321
1322        if (nr_taken == 0)
1323                return 0;
1324
1325        nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1326                                        &nr_dirty, &nr_writeback, false);
1327
1328        spin_lock_irq(&zone->lru_lock);
1329
1330        reclaim_stat->recent_scanned[file] += nr_taken;
1331
1332        if (global_reclaim(sc)) {
1333                if (current_is_kswapd())
1334                        __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1335                                               nr_reclaimed);
1336                else
1337                        __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1338                                               nr_reclaimed);
1339        }
1340
1341        putback_inactive_pages(lruvec, &page_list);
1342
1343        __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1344
1345        spin_unlock_irq(&zone->lru_lock);
1346
1347        free_hot_cold_page_list(&page_list, 1);
1348
1349        /*
1350         * If reclaim is isolating dirty pages under writeback, it implies
1351         * that the long-lived page allocation rate is exceeding the page
1352         * laundering rate. Either the global limits are not being effective
1353         * at throttling processes due to the page distribution throughout
1354         * zones or there is heavy usage of a slow backing device. The
1355         * only option is to throttle from reclaim context which is not ideal
1356         * as there is no guarantee the dirtying process is throttled in the
1357         * same way balance_dirty_pages() manages.
1358         *
1359         * This scales the number of dirty pages that must be under writeback
1360         * before throttling depending on priority. It is a simple backoff
1361         * function that has the most effect in the range DEF_PRIORITY to
1362         * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1363         * in trouble and reclaim is considered to be in trouble.
1364         *
1365         * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1366         * DEF_PRIORITY-1  50% must be PageWriteback
1367         * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1368         * ...
1369         * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1370         *                     isolated page is PageWriteback
1371         */
1372        if (nr_writeback && nr_writeback >=
1373                        (nr_taken >> (DEF_PRIORITY - sc->priority)))
1374                wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1375
1376        trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1377                zone_idx(zone),
1378                nr_scanned, nr_reclaimed,
1379                sc->priority,
1380                trace_shrink_flags(file));
1381        return nr_reclaimed;
1382}
1383
1384/*
1385 * This moves pages from the active list to the inactive list.
1386 *
1387 * We move them the other way if the page is referenced by one or more
1388 * processes, from rmap.
1389 *
1390 * If the pages are mostly unmapped, the processing is fast and it is
1391 * appropriate to hold zone->lru_lock across the whole operation.  But if
1392 * the pages are mapped, the processing is slow (page_referenced()) so we
1393 * should drop zone->lru_lock around each page.  It's impossible to balance
1394 * this, so instead we remove the pages from the LRU while processing them.
1395 * It is safe to rely on PG_active against the non-LRU pages in here because
1396 * nobody will play with that bit on a non-LRU page.
1397 *
1398 * The downside is that we have to touch page->_count against each page.
1399 * But we had to alter page->flags anyway.
1400 */
1401
1402static void move_active_pages_to_lru(struct lruvec *lruvec,
1403                                     struct list_head *list,
1404                                     struct list_head *pages_to_free,
1405                                     enum lru_list lru)
1406{
1407        struct zone *zone = lruvec_zone(lruvec);
1408        unsigned long pgmoved = 0;
1409        struct page *page;
1410        int nr_pages;
1411
1412        while (!list_empty(list)) {
1413                page = lru_to_page(list);
1414                lruvec = mem_cgroup_page_lruvec(page, zone);
1415
1416                VM_BUG_ON(PageLRU(page));
1417                SetPageLRU(page);
1418
1419                nr_pages = hpage_nr_pages(page);
1420                mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1421                list_move(&page->lru, &lruvec->lists[lru]);
1422                pgmoved += nr_pages;
1423
1424                if (put_page_testzero(page)) {
1425                        __ClearPageLRU(page);
1426                        __ClearPageActive(page);
1427                        del_page_from_lru_list(page, lruvec, lru);
1428
1429                        if (unlikely(PageCompound(page))) {
1430                                spin_unlock_irq(&zone->lru_lock);
1431                                (*get_compound_page_dtor(page))(page);
1432                                spin_lock_irq(&zone->lru_lock);
1433                        } else
1434                                list_add(&page->lru, pages_to_free);
1435                }
1436        }
1437        __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1438        if (!is_active_lru(lru))
1439                __count_vm_events(PGDEACTIVATE, pgmoved);
1440}
1441
1442static void shrink_active_list(unsigned long nr_to_scan,
1443                               struct lruvec *lruvec,
1444                               struct scan_control *sc,
1445                               enum lru_list lru)
1446{
1447        unsigned long nr_taken;
1448        unsigned long nr_scanned;
1449        unsigned long vm_flags;
1450        LIST_HEAD(l_hold);      /* The pages which were snipped off */
1451        LIST_HEAD(l_active);
1452        LIST_HEAD(l_inactive);
1453        struct page *page;
1454        struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1455        unsigned long nr_rotated = 0;
1456        isolate_mode_t isolate_mode = 0;
1457        int file = is_file_lru(lru);
1458        struct zone *zone = lruvec_zone(lruvec);
1459
1460        lru_add_drain();
1461
1462        if (!sc->may_unmap)
1463                isolate_mode |= ISOLATE_UNMAPPED;
1464        if (!sc->may_writepage)
1465                isolate_mode |= ISOLATE_CLEAN;
1466
1467        spin_lock_irq(&zone->lru_lock);
1468
1469        nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1470                                     &nr_scanned, sc, isolate_mode, lru);
1471        if (global_reclaim(sc))
1472                zone->pages_scanned += nr_scanned;
1473
1474        reclaim_stat->recent_scanned[file] += nr_taken;
1475
1476        __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1477        __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1478        __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1479        spin_unlock_irq(&zone->lru_lock);
1480
1481        while (!list_empty(&l_hold)) {
1482                cond_resched();
1483                page = lru_to_page(&l_hold);
1484                list_del(&page->lru);
1485
1486                if (unlikely(!page_evictable(page))) {
1487                        putback_lru_page(page);
1488                        continue;
1489                }
1490
1491                if (unlikely(buffer_heads_over_limit)) {
1492                        if (page_has_private(page) && trylock_page(page)) {
1493                                if (page_has_private(page))
1494                                        try_to_release_page(page, 0);
1495                                unlock_page(page);
1496                        }
1497                }
1498
1499                if (page_referenced(page, 0, sc->target_mem_cgroup,
1500                                    &vm_flags)) {
1501                        nr_rotated += hpage_nr_pages(page);
1502                        /*
1503                         * Identify referenced, file-backed active pages and
1504                         * give them one more trip around the active list. So
1505                         * that executable code get better chances to stay in
1506                         * memory under moderate memory pressure.  Anon pages
1507                         * are not likely to be evicted by use-once streaming
1508                         * IO, plus JVM can create lots of anon VM_EXEC pages,
1509                         * so we ignore them here.
1510                         */
1511                        if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1512                                list_add(&page->lru, &l_active);
1513                                continue;
1514                        }
1515                }
1516
1517                ClearPageActive(page);  /* we are de-activating */
1518                list_add(&page->lru, &l_inactive);
1519        }
1520
1521        /*
1522         * Move pages back to the lru list.
1523         */
1524        spin_lock_irq(&zone->lru_lock);
1525        /*
1526         * Count referenced pages from currently used mappings as rotated,
1527         * even though only some of them are actually re-activated.  This
1528         * helps balance scan pressure between file and anonymous pages in
1529         * get_scan_ratio.
1530         */
1531        reclaim_stat->recent_rotated[file] += nr_rotated;
1532
1533        move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1534        move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1535        __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1536        spin_unlock_irq(&zone->lru_lock);
1537
1538        free_hot_cold_page_list(&l_hold, 1);
1539}
1540
1541#ifdef CONFIG_SWAP
1542static int inactive_anon_is_low_global(struct zone *zone)
1543{
1544        unsigned long active, inactive;
1545
1546        active = zone_page_state(zone, NR_ACTIVE_ANON);
1547        inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1548
1549        if (inactive * zone->inactive_ratio < active)
1550                return 1;
1551
1552        return 0;
1553}
1554
1555/**
1556 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1557 * @lruvec: LRU vector to check
1558 *
1559 * Returns true if the zone does not have enough inactive anon pages,
1560 * meaning some active anon pages need to be deactivated.
1561 */
1562static int inactive_anon_is_low(struct lruvec *lruvec)
1563{
1564        /*
1565         * If we don't have swap space, anonymous page deactivation
1566         * is pointless.
1567         */
1568        if (!total_swap_pages)
1569                return 0;
1570
1571        if (!mem_cgroup_disabled())
1572                return mem_cgroup_inactive_anon_is_low(lruvec);
1573
1574        return inactive_anon_is_low_global(lruvec_zone(lruvec));
1575}
1576#else
1577static inline int inactive_anon_is_low(struct lruvec *lruvec)
1578{
1579        return 0;
1580}
1581#endif
1582
1583/**
1584 * inactive_file_is_low - check if file pages need to be deactivated
1585 * @lruvec: LRU vector to check
1586 *
1587 * When the system is doing streaming IO, memory pressure here
1588 * ensures that active file pages get deactivated, until more
1589 * than half of the file pages are on the inactive list.
1590 *
1591 * Once we get to that situation, protect the system's working
1592 * set from being evicted by disabling active file page aging.
1593 *
1594 * This uses a different ratio than the anonymous pages, because
1595 * the page cache uses a use-once replacement algorithm.
1596 */
1597static int inactive_file_is_low(struct lruvec *lruvec)
1598{
1599        unsigned long inactive;
1600        unsigned long active;
1601
1602        inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1603        active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1604
1605        return active > inactive;
1606}
1607
1608static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1609{
1610        if (is_file_lru(lru))
1611                return inactive_file_is_low(lruvec);
1612        else
1613                return inactive_anon_is_low(lruvec);
1614}
1615
1616static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1617                                 struct lruvec *lruvec, struct scan_control *sc)
1618{
1619        if (is_active_lru(lru)) {
1620                if (inactive_list_is_low(lruvec, lru))
1621                        shrink_active_list(nr_to_scan, lruvec, sc, lru);
1622                return 0;
1623        }
1624
1625        return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1626}
1627
1628static int vmscan_swappiness(struct scan_control *sc)
1629{
1630        if (global_reclaim(sc))
1631                return vm_swappiness;
1632        return mem_cgroup_swappiness(sc->target_mem_cgroup);
1633}
1634
1635enum scan_balance {
1636        SCAN_EQUAL,
1637        SCAN_FRACT,
1638        SCAN_ANON,
1639        SCAN_FILE,
1640};
1641
1642/*
1643 * Determine how aggressively the anon and file LRU lists should be
1644 * scanned.  The relative value of each set of LRU lists is determined
1645 * by looking at the fraction of the pages scanned we did rotate back
1646 * onto the active list instead of evict.
1647 *
1648 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1649 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1650 */
1651static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1652                           unsigned long *nr)
1653{
1654        struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1655        u64 fraction[2];
1656        u64 denominator = 0;    /* gcc */
1657        struct zone *zone = lruvec_zone(lruvec);
1658        unsigned long anon_prio, file_prio;
1659        enum scan_balance scan_balance;
1660        unsigned long anon, file, free;
1661        bool force_scan = false;
1662        unsigned long ap, fp;
1663        enum lru_list lru;
1664
1665        /*
1666         * If the zone or memcg is small, nr[l] can be 0.  This
1667         * results in no scanning on this priority and a potential
1668         * priority drop.  Global direct reclaim can go to the next
1669         * zone and tends to have no problems. Global kswapd is for
1670         * zone balancing and it needs to scan a minimum amount. When
1671         * reclaiming for a memcg, a priority drop can cause high
1672         * latencies, so it's better to scan a minimum amount there as
1673         * well.
1674         */
1675        if (current_is_kswapd() && zone->all_unreclaimable)
1676                force_scan = true;
1677        if (!global_reclaim(sc))
1678                force_scan = true;
1679
1680        /* If we have no swap space, do not bother scanning anon pages. */
1681        if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1682                scan_balance = SCAN_FILE;
1683                goto out;
1684        }
1685
1686        /*
1687         * Global reclaim will swap to prevent OOM even with no
1688         * swappiness, but memcg users want to use this knob to
1689         * disable swapping for individual groups completely when
1690         * using the memory controller's swap limit feature would be
1691         * too expensive.
1692         */
1693        if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
1694                scan_balance = SCAN_FILE;
1695                goto out;
1696        }
1697
1698        /*
1699         * Do not apply any pressure balancing cleverness when the
1700         * system is close to OOM, scan both anon and file equally
1701         * (unless the swappiness setting disagrees with swapping).
1702         */
1703        if (!sc->priority && vmscan_swappiness(sc)) {
1704                scan_balance = SCAN_EQUAL;
1705                goto out;
1706        }
1707
1708        anon  = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1709                get_lru_size(lruvec, LRU_INACTIVE_ANON);
1710        file  = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1711                get_lru_size(lruvec, LRU_INACTIVE_FILE);
1712
1713        /*
1714         * If it's foreseeable that reclaiming the file cache won't be
1715         * enough to get the zone back into a desirable shape, we have
1716         * to swap.  Better start now and leave the - probably heavily
1717         * thrashing - remaining file pages alone.
1718         */
1719        if (global_reclaim(sc)) {
1720                free = zone_page_state(zone, NR_FREE_PAGES);
1721                if (unlikely(file + free <= high_wmark_pages(zone))) {
1722                        scan_balance = SCAN_ANON;
1723                        goto out;
1724                }
1725        }
1726
1727        /*
1728         * There is enough inactive page cache, do not reclaim
1729         * anything from the anonymous working set right now.
1730         */
1731        if (!inactive_file_is_low(lruvec)) {
1732                scan_balance = SCAN_FILE;
1733                goto out;
1734        }
1735
1736        scan_balance = SCAN_FRACT;
1737
1738        /*
1739         * With swappiness at 100, anonymous and file have the same priority.
1740         * This scanning priority is essentially the inverse of IO cost.
1741         */
1742        anon_prio = vmscan_swappiness(sc);
1743        file_prio = 200 - anon_prio;
1744
1745        /*
1746         * OK, so we have swap space and a fair amount of page cache
1747         * pages.  We use the recently rotated / recently scanned
1748         * ratios to determine how valuable each cache is.
1749         *
1750         * Because workloads change over time (and to avoid overflow)
1751         * we keep these statistics as a floating average, which ends
1752         * up weighing recent references more than old ones.
1753         *
1754         * anon in [0], file in [1]
1755         */
1756        spin_lock_irq(&zone->lru_lock);
1757        if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1758                reclaim_stat->recent_scanned[0] /= 2;
1759                reclaim_stat->recent_rotated[0] /= 2;
1760        }
1761
1762        if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1763                reclaim_stat->recent_scanned[1] /= 2;
1764                reclaim_stat->recent_rotated[1] /= 2;
1765        }
1766
1767        /*
1768         * The amount of pressure on anon vs file pages is inversely
1769         * proportional to the fraction of recently scanned pages on
1770         * each list that were recently referenced and in active use.
1771         */
1772        ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1773        ap /= reclaim_stat->recent_rotated[0] + 1;
1774
1775        fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1776        fp /= reclaim_stat->recent_rotated[1] + 1;
1777        spin_unlock_irq(&zone->lru_lock);
1778
1779        fraction[0] = ap;
1780        fraction[1] = fp;
1781        denominator = ap + fp + 1;
1782out:
1783        for_each_evictable_lru(lru) {
1784                int file = is_file_lru(lru);
1785                unsigned long size;
1786                unsigned long scan;
1787
1788                size = get_lru_size(lruvec, lru);
1789                scan = size >> sc->priority;
1790
1791                if (!scan && force_scan)
1792                        scan = min(size, SWAP_CLUSTER_MAX);
1793
1794                switch (scan_balance) {
1795                case SCAN_EQUAL:
1796                        /* Scan lists relative to size */
1797                        break;
1798                case SCAN_FRACT:
1799                        /*
1800                         * Scan types proportional to swappiness and
1801                         * their relative recent reclaim efficiency.
1802                         */
1803                        scan = div64_u64(scan * fraction[file], denominator);
1804                        break;
1805                case SCAN_FILE:
1806                case SCAN_ANON:
1807                        /* Scan one type exclusively */
1808                        if ((scan_balance == SCAN_FILE) != file)
1809                                scan = 0;
1810                        break;
1811                default:
1812                        /* Look ma, no brain */
1813                        BUG();
1814                }
1815                nr[lru] = scan;
1816        }
1817}
1818
1819/*
1820 * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1821 */
1822static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1823{
1824        unsigned long nr[NR_LRU_LISTS];
1825        unsigned long nr_to_scan;
1826        enum lru_list lru;
1827        unsigned long nr_reclaimed = 0;
1828        unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1829        struct blk_plug plug;
1830
1831        get_scan_count(lruvec, sc, nr);
1832
1833        blk_start_plug(&plug);
1834        while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1835                                        nr[LRU_INACTIVE_FILE]) {
1836                for_each_evictable_lru(lru) {
1837                        if (nr[lru]) {
1838                                nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
1839                                nr[lru] -= nr_to_scan;
1840
1841                                nr_reclaimed += shrink_list(lru, nr_to_scan,
1842                                                            lruvec, sc);
1843                        }
1844                }
1845                /*
1846                 * On large memory systems, scan >> priority can become
1847                 * really large. This is fine for the starting priority;
1848                 * we want to put equal scanning pressure on each zone.
1849                 * However, if the VM has a harder time of freeing pages,
1850                 * with multiple processes reclaiming pages, the total
1851                 * freeing target can get unreasonably large.
1852                 */
1853                if (nr_reclaimed >= nr_to_reclaim &&
1854                    sc->priority < DEF_PRIORITY)
1855                        break;
1856        }
1857        blk_finish_plug(&plug);
1858        sc->nr_reclaimed += nr_reclaimed;
1859
1860        /*
1861         * Even if we did not try to evict anon pages at all, we want to
1862         * rebalance the anon lru active/inactive ratio.
1863         */
1864        if (inactive_anon_is_low(lruvec))
1865                shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
1866                                   sc, LRU_ACTIVE_ANON);
1867
1868        throttle_vm_writeout(sc->gfp_mask);
1869}
1870
1871/* Use reclaim/compaction for costly allocs or under memory pressure */
1872static bool in_reclaim_compaction(struct scan_control *sc)
1873{
1874        if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
1875                        (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
1876                         sc->priority < DEF_PRIORITY - 2))
1877                return true;
1878
1879        return false;
1880}
1881
1882/*
1883 * Reclaim/compaction is used for high-order allocation requests. It reclaims
1884 * order-0 pages before compacting the zone. should_continue_reclaim() returns
1885 * true if more pages should be reclaimed such that when the page allocator
1886 * calls try_to_compact_zone() that it will have enough free pages to succeed.
1887 * It will give up earlier than that if there is difficulty reclaiming pages.
1888 */
1889static inline bool should_continue_reclaim(struct zone *zone,
1890                                        unsigned long nr_reclaimed,
1891                                        unsigned long nr_scanned,
1892                                        struct scan_control *sc)
1893{
1894        unsigned long pages_for_compaction;
1895        unsigned long inactive_lru_pages;
1896
1897        /* If not in reclaim/compaction mode, stop */
1898        if (!in_reclaim_compaction(sc))
1899                return false;
1900
1901        /* Consider stopping depending on scan and reclaim activity */
1902        if (sc->gfp_mask & __GFP_REPEAT) {
1903                /*
1904                 * For __GFP_REPEAT allocations, stop reclaiming if the
1905                 * full LRU list has been scanned and we are still failing
1906                 * to reclaim pages. This full LRU scan is potentially
1907                 * expensive but a __GFP_REPEAT caller really wants to succeed
1908                 */
1909                if (!nr_reclaimed && !nr_scanned)
1910                        return false;
1911        } else {
1912                /*
1913                 * For non-__GFP_REPEAT allocations which can presumably
1914                 * fail without consequence, stop if we failed to reclaim
1915                 * any pages from the last SWAP_CLUSTER_MAX number of
1916                 * pages that were scanned. This will return to the
1917                 * caller faster at the risk reclaim/compaction and
1918                 * the resulting allocation attempt fails
1919                 */
1920                if (!nr_reclaimed)
1921                        return false;
1922        }
1923
1924        /*
1925         * If we have not reclaimed enough pages for compaction and the
1926         * inactive lists are large enough, continue reclaiming
1927         */
1928        pages_for_compaction = (2UL << sc->order);
1929        inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
1930        if (get_nr_swap_pages() > 0)
1931                inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
1932        if (sc->nr_reclaimed < pages_for_compaction &&
1933                        inactive_lru_pages > pages_for_compaction)
1934                return true;
1935
1936        /* If compaction would go ahead or the allocation would succeed, stop */
1937        switch (compaction_suitable(zone, sc->order)) {
1938        case COMPACT_PARTIAL:
1939        case COMPACT_CONTINUE:
1940                return false;
1941        default:
1942                return true;
1943        }
1944}
1945
1946static void shrink_zone(struct zone *zone, struct scan_control *sc)
1947{
1948        unsigned long nr_reclaimed, nr_scanned;
1949
1950        do {
1951                struct mem_cgroup *root = sc->target_mem_cgroup;
1952                struct mem_cgroup_reclaim_cookie reclaim = {
1953                        .zone = zone,
1954                        .priority = sc->priority,
1955                };
1956                struct mem_cgroup *memcg;
1957
1958                nr_reclaimed = sc->nr_reclaimed;
1959                nr_scanned = sc->nr_scanned;
1960
1961                memcg = mem_cgroup_iter(root, NULL, &reclaim);
1962                do {
1963                        struct lruvec *lruvec;
1964
1965                        lruvec = mem_cgroup_zone_lruvec(zone, memcg);
1966
1967                        shrink_lruvec(lruvec, sc);
1968
1969                        /*
1970                         * Direct reclaim and kswapd have to scan all memory
1971                         * cgroups to fulfill the overall scan target for the
1972                         * zone.
1973                         *
1974                         * Limit reclaim, on the other hand, only cares about
1975                         * nr_to_reclaim pages to be reclaimed and it will
1976                         * retry with decreasing priority if one round over the
1977                         * whole hierarchy is not sufficient.
1978                         */
1979                        if (!global_reclaim(sc) &&
1980                                        sc->nr_reclaimed >= sc->nr_to_reclaim) {
1981                                mem_cgroup_iter_break(root, memcg);
1982                                break;
1983                        }
1984                        memcg = mem_cgroup_iter(root, memcg, &reclaim);
1985                } while (memcg);
1986
1987                vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
1988                           sc->nr_scanned - nr_scanned,
1989                           sc->nr_reclaimed - nr_reclaimed);
1990
1991        } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
1992                                         sc->nr_scanned - nr_scanned, sc));
1993}
1994
1995/* Returns true if compaction should go ahead for a high-order request */
1996static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
1997{
1998        unsigned long balance_gap, watermark;
1999        bool watermark_ok;
2000
2001        /* Do not consider compaction for orders reclaim is meant to satisfy */
2002        if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2003                return false;
2004
2005        /*
2006         * Compaction takes time to run and there are potentially other
2007         * callers using the pages just freed. Continue reclaiming until
2008         * there is a buffer of free pages available to give compaction
2009         * a reasonable chance of completing and allocating the page
2010         */
2011        balance_gap = min(low_wmark_pages(zone),
2012                (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2013                        KSWAPD_ZONE_BALANCE_GAP_RATIO);
2014        watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2015        watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2016
2017        /*
2018         * If compaction is deferred, reclaim up to a point where
2019         * compaction will have a chance of success when re-enabled
2020         */
2021        if (compaction_deferred(zone, sc->order))
2022                return watermark_ok;
2023
2024        /* If compaction is not ready to start, keep reclaiming */
2025        if (!compaction_suitable(zone, sc->order))
2026                return false;
2027
2028        return watermark_ok;
2029}
2030
2031/*
2032 * This is the direct reclaim path, for page-allocating processes.  We only
2033 * try to reclaim pages from zones which will satisfy the caller's allocation
2034 * request.
2035 *
2036 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2037 * Because:
2038 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2039 *    allocation or
2040 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2041 *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2042 *    zone defense algorithm.
2043 *
2044 * If a zone is deemed to be full of pinned pages then just give it a light
2045 * scan then give up on it.
2046 *
2047 * This function returns true if a zone is being reclaimed for a costly
2048 * high-order allocation and compaction is ready to begin. This indicates to
2049 * the caller that it should consider retrying the allocation instead of
2050 * further reclaim.
2051 */
2052static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2053{
2054        struct zoneref *z;
2055        struct zone *zone;
2056        unsigned long nr_soft_reclaimed;
2057        unsigned long nr_soft_scanned;
2058        bool aborted_reclaim = false;
2059
2060        /*
2061         * If the number of buffer_heads in the machine exceeds the maximum
2062         * allowed level, force direct reclaim to scan the highmem zone as
2063         * highmem pages could be pinning lowmem pages storing buffer_heads
2064         */
2065        if (buffer_heads_over_limit)
2066                sc->gfp_mask |= __GFP_HIGHMEM;
2067
2068        for_each_zone_zonelist_nodemask(zone, z, zonelist,
2069                                        gfp_zone(sc->gfp_mask), sc->nodemask) {
2070                if (!populated_zone(zone))
2071                        continue;
2072                /*
2073                 * Take care memory controller reclaiming has small influence
2074                 * to global LRU.
2075                 */
2076                if (global_reclaim(sc)) {
2077                        if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2078                                continue;
2079                        if (zone->all_unreclaimable &&
2080                                        sc->priority != DEF_PRIORITY)
2081                                continue;       /* Let kswapd poll it */
2082                        if (IS_ENABLED(CONFIG_COMPACTION)) {
2083                                /*
2084                                 * If we already have plenty of memory free for
2085                                 * compaction in this zone, don't free any more.
2086                                 * Even though compaction is invoked for any
2087                                 * non-zero order, only frequent costly order
2088                                 * reclamation is disruptive enough to become a
2089                                 * noticeable problem, like transparent huge
2090                                 * page allocations.
2091                                 */
2092                                if (compaction_ready(zone, sc)) {
2093                                        aborted_reclaim = true;
2094                                        continue;
2095                                }
2096                        }
2097                        /*
2098                         * This steals pages from memory cgroups over softlimit
2099                         * and returns the number of reclaimed pages and
2100                         * scanned pages. This works for global memory pressure
2101                         * and balancing, not for a memcg's limit.
2102                         */
2103                        nr_soft_scanned = 0;
2104                        nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2105                                                sc->order, sc->gfp_mask,
2106                                                &nr_soft_scanned);
2107                        sc->nr_reclaimed += nr_soft_reclaimed;
2108                        sc->nr_scanned += nr_soft_scanned;
2109                        /* need some check for avoid more shrink_zone() */
2110                }
2111
2112                shrink_zone(zone, sc);
2113        }
2114
2115        return aborted_reclaim;
2116}
2117
2118static bool zone_reclaimable(struct zone *zone)
2119{
2120        return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2121}
2122
2123/* All zones in zonelist are unreclaimable? */
2124static bool all_unreclaimable(struct zonelist *zonelist,
2125                struct scan_control *sc)
2126{
2127        struct zoneref *z;
2128        struct zone *zone;
2129
2130        for_each_zone_zonelist_nodemask(zone, z, zonelist,
2131                        gfp_zone(sc->gfp_mask), sc->nodemask) {
2132                if (!populated_zone(zone))
2133                        continue;
2134                if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2135                        continue;
2136                if (!zone->all_unreclaimable)
2137                        return false;
2138        }
2139
2140        return true;
2141}
2142
2143/*
2144 * This is the main entry point to direct page reclaim.
2145 *
2146 * If a full scan of the inactive list fails to free enough memory then we
2147 * are "out of memory" and something needs to be killed.
2148 *
2149 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2150 * high - the zone may be full of dirty or under-writeback pages, which this
2151 * caller can't do much about.  We kick the writeback threads and take explicit
2152 * naps in the hope that some of these pages can be written.  But if the
2153 * allocating task holds filesystem locks which prevent writeout this might not
2154 * work, and the allocation attempt will fail.
2155 *
2156 * returns:     0, if no pages reclaimed
2157 *              else, the number of pages reclaimed
2158 */
2159static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2160                                        struct scan_control *sc,
2161                                        struct shrink_control *shrink)
2162{
2163        unsigned long total_scanned = 0;
2164        struct reclaim_state *reclaim_state = current->reclaim_state;
2165        struct zoneref *z;
2166        struct zone *zone;
2167        unsigned long writeback_threshold;
2168        bool aborted_reclaim;
2169
2170        delayacct_freepages_start();
2171
2172        if (global_reclaim(sc))
2173                count_vm_event(ALLOCSTALL);
2174
2175        do {
2176                vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2177                                sc->priority);
2178                sc->nr_scanned = 0;
2179                aborted_reclaim = shrink_zones(zonelist, sc);
2180
2181                /*
2182                 * Don't shrink slabs when reclaiming memory from
2183                 * over limit cgroups
2184                 */
2185                if (global_reclaim(sc)) {
2186                        unsigned long lru_pages = 0;
2187                        for_each_zone_zonelist(zone, z, zonelist,
2188                                        gfp_zone(sc->gfp_mask)) {
2189                                if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2190                                        continue;
2191
2192                                lru_pages += zone_reclaimable_pages(zone);
2193                        }
2194
2195                        shrink_slab(shrink, sc->nr_scanned, lru_pages);
2196                        if (reclaim_state) {
2197                                sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2198                                reclaim_state->reclaimed_slab = 0;
2199                        }
2200                }
2201                total_scanned += sc->nr_scanned;
2202                if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2203                        goto out;
2204
2205                /*
2206                 * If we're getting trouble reclaiming, start doing
2207                 * writepage even in laptop mode.
2208                 */
2209                if (sc->priority < DEF_PRIORITY - 2)
2210                        sc->may_writepage = 1;
2211
2212                /*
2213                 * Try to write back as many pages as we just scanned.  This
2214                 * tends to cause slow streaming writers to write data to the
2215                 * disk smoothly, at the dirtying rate, which is nice.   But
2216                 * that's undesirable in laptop mode, where we *want* lumpy
2217                 * writeout.  So in laptop mode, write out the whole world.
2218                 */
2219                writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2220                if (total_scanned > writeback_threshold) {
2221                        wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2222                                                WB_REASON_TRY_TO_FREE_PAGES);
2223                        sc->may_writepage = 1;
2224                }
2225
2226                /* Take a nap, wait for some writeback to complete */
2227                if (!sc->hibernation_mode && sc->nr_scanned &&
2228                    sc->priority < DEF_PRIORITY - 2) {
2229                        struct zone *preferred_zone;
2230
2231                        first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2232                                                &cpuset_current_mems_allowed,
2233                                                &preferred_zone);
2234                        wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2235                }
2236        } while (--sc->priority >= 0);
2237
2238out:
2239        delayacct_freepages_end();
2240
2241        if (sc->nr_reclaimed)
2242                return sc->nr_reclaimed;
2243
2244        /*
2245         * As hibernation is going on, kswapd is freezed so that it can't mark
2246         * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2247         * check.
2248         */
2249        if (oom_killer_disabled)
2250                return 0;
2251
2252        /* Aborted reclaim to try compaction? don't OOM, then */
2253        if (aborted_reclaim)
2254                return 1;
2255
2256        /* top priority shrink_zones still had more to do? don't OOM, then */
2257        if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2258                return 1;
2259
2260        return 0;
2261}
2262
2263static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2264{
2265        struct zone *zone;
2266        unsigned long pfmemalloc_reserve = 0;
2267        unsigned long free_pages = 0;
2268        int i;
2269        bool wmark_ok;
2270
2271        for (i = 0; i <= ZONE_NORMAL; i++) {
2272                zone = &pgdat->node_zones[i];
2273                pfmemalloc_reserve += min_wmark_pages(zone);
2274                free_pages += zone_page_state(zone, NR_FREE_PAGES);
2275        }
2276
2277        wmark_ok = free_pages > pfmemalloc_reserve / 2;
2278
2279        /* kswapd must be awake if processes are being throttled */
2280        if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2281                pgdat->classzone_idx = min(pgdat->classzone_idx,
2282                                                (enum zone_type)ZONE_NORMAL);
2283                wake_up_interruptible(&pgdat->kswapd_wait);
2284        }
2285
2286        return wmark_ok;
2287}
2288
2289/*
2290 * Throttle direct reclaimers if backing storage is backed by the network
2291 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2292 * depleted. kswapd will continue to make progress and wake the processes
2293 * when the low watermark is reached.
2294 *
2295 * Returns true if a fatal signal was delivered during throttling. If this
2296 * happens, the page allocator should not consider triggering the OOM killer.
2297 */
2298static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2299                                        nodemask_t *nodemask)
2300{
2301        struct zone *zone;
2302        int high_zoneidx = gfp_zone(gfp_mask);
2303        pg_data_t *pgdat;
2304
2305        /*
2306         * Kernel threads should not be throttled as they may be indirectly
2307         * responsible for cleaning pages necessary for reclaim to make forward
2308         * progress. kjournald for example may enter direct reclaim while
2309         * committing a transaction where throttling it could forcing other
2310         * processes to block on log_wait_commit().
2311         */
2312        if (current->flags & PF_KTHREAD)
2313                goto out;
2314
2315        /*
2316         * If a fatal signal is pending, this process should not throttle.
2317         * It should return quickly so it can exit and free its memory
2318         */
2319        if (fatal_signal_pending(current))
2320                goto out;
2321
2322        /* Check if the pfmemalloc reserves are ok */
2323        first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2324        pgdat = zone->zone_pgdat;
2325        if (pfmemalloc_watermark_ok(pgdat))
2326                goto out;
2327
2328        /* Account for the throttling */
2329        count_vm_event(PGSCAN_DIRECT_THROTTLE);
2330
2331        /*
2332         * If the caller cannot enter the filesystem, it's possible that it
2333         * is due to the caller holding an FS lock or performing a journal
2334         * transaction in the case of a filesystem like ext[3|4]. In this case,
2335         * it is not safe to block on pfmemalloc_wait as kswapd could be
2336         * blocked waiting on the same lock. Instead, throttle for up to a
2337         * second before continuing.
2338         */
2339        if (!(gfp_mask & __GFP_FS)) {
2340                wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2341                        pfmemalloc_watermark_ok(pgdat), HZ);
2342
2343                goto check_pending;
2344        }
2345
2346        /* Throttle until kswapd wakes the process */
2347        wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2348                pfmemalloc_watermark_ok(pgdat));
2349
2350check_pending:
2351        if (fatal_signal_pending(current))
2352                return true;
2353
2354out:
2355        return false;
2356}
2357
2358unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2359                                gfp_t gfp_mask, nodemask_t *nodemask)
2360{
2361        unsigned long nr_reclaimed;
2362        struct scan_control sc = {
2363                .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2364                .may_writepage = !laptop_mode,
2365                .nr_to_reclaim = SWAP_CLUSTER_MAX,
2366                .may_unmap = 1,
2367                .may_swap = 1,
2368                .order = order,
2369                .priority = DEF_PRIORITY,
2370                .target_mem_cgroup = NULL,
2371                .nodemask = nodemask,
2372        };
2373        struct shrink_control shrink = {
2374                .gfp_mask = sc.gfp_mask,
2375        };
2376
2377        /*
2378         * Do not enter reclaim if fatal signal was delivered while throttled.
2379         * 1 is returned so that the page allocator does not OOM kill at this
2380         * point.
2381         */
2382        if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2383                return 1;
2384
2385        trace_mm_vmscan_direct_reclaim_begin(order,
2386                                sc.may_writepage,
2387                                gfp_mask);
2388
2389        nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2390
2391        trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2392
2393        return nr_reclaimed;
2394}
2395
2396#ifdef CONFIG_MEMCG
2397
2398unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2399                                                gfp_t gfp_mask, bool noswap,
2400                                                struct zone *zone,
2401                                                unsigned long *nr_scanned)
2402{
2403        struct scan_control sc = {
2404                .nr_scanned = 0,
2405                .nr_to_reclaim = SWAP_CLUSTER_MAX,
2406                .may_writepage = !laptop_mode,
2407                .may_unmap = 1,
2408                .may_swap = !noswap,
2409                .order = 0,
2410                .priority = 0,
2411                .target_mem_cgroup = memcg,
2412        };
2413        struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2414
2415        sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2416                        (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2417
2418        trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2419                                                      sc.may_writepage,
2420                                                      sc.gfp_mask);
2421
2422        /*
2423         * NOTE: Although we can get the priority field, using it
2424         * here is not a good idea, since it limits the pages we can scan.
2425         * if we don't reclaim here, the shrink_zone from balance_pgdat
2426         * will pick up pages from other mem cgroup's as well. We hack
2427         * the priority and make it zero.
2428         */
2429        shrink_lruvec(lruvec, &sc);
2430
2431        trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2432
2433        *nr_scanned = sc.nr_scanned;
2434        return sc.nr_reclaimed;
2435}
2436
2437unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2438                                           gfp_t gfp_mask,
2439                                           bool noswap)
2440{
2441        struct zonelist *zonelist;
2442        unsigned long nr_reclaimed;
2443        int nid;
2444        struct scan_control sc = {
2445                .may_writepage = !laptop_mode,
2446                .may_unmap = 1,
2447                .may_swap = !noswap,
2448                .nr_to_reclaim = SWAP_CLUSTER_MAX,
2449                .order = 0,
2450                .priority = DEF_PRIORITY,
2451                .target_mem_cgroup = memcg,
2452                .nodemask = NULL, /* we don't care the placement */
2453                .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2454                                (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2455        };
2456        struct shrink_control shrink = {
2457                .gfp_mask = sc.gfp_mask,
2458        };
2459
2460        /*
2461         * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2462         * take care of from where we get pages. So the node where we start the
2463         * scan does not need to be the current node.
2464         */
2465        nid = mem_cgroup_select_victim_node(memcg);
2466
2467        zonelist = NODE_DATA(nid)->node_zonelists;
2468
2469        trace_mm_vmscan_memcg_reclaim_begin(0,
2470                                            sc.may_writepage,
2471                                            sc.gfp_mask);
2472
2473        nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2474
2475        trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2476
2477        return nr_reclaimed;
2478}
2479#endif
2480
2481static void age_active_anon(struct zone *zone, struct scan_control *sc)
2482{
2483        struct mem_cgroup *memcg;
2484
2485        if (!total_swap_pages)
2486                return;
2487
2488        memcg = mem_cgroup_iter(NULL, NULL, NULL);
2489        do {
2490                struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2491
2492                if (inactive_anon_is_low(lruvec))
2493                        shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2494                                           sc, LRU_ACTIVE_ANON);
2495
2496                memcg = mem_cgroup_iter(NULL, memcg, NULL);
2497        } while (memcg);
2498}
2499
2500static bool zone_balanced(struct zone *zone, int order,
2501                          unsigned long balance_gap, int classzone_idx)
2502{
2503        if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2504                                    balance_gap, classzone_idx, 0))
2505                return false;
2506
2507        if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2508            !compaction_suitable(zone, order))
2509                return false;
2510
2511        return true;
2512}
2513
2514/*
2515 * pgdat_balanced() is used when checking if a node is balanced.
2516 *
2517 * For order-0, all zones must be balanced!
2518 *
2519 * For high-order allocations only zones that meet watermarks and are in a
2520 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2521 * total of balanced pages must be at least 25% of the zones allowed by
2522 * classzone_idx for the node to be considered balanced. Forcing all zones to
2523 * be balanced for high orders can cause excessive reclaim when there are
2524 * imbalanced zones.
2525 * The choice of 25% is due to
2526 *   o a 16M DMA zone that is balanced will not balance a zone on any
2527 *     reasonable sized machine
2528 *   o On all other machines, the top zone must be at least a reasonable
2529 *     percentage of the middle zones. For example, on 32-bit x86, highmem
2530 *     would need to be at least 256M for it to be balance a whole node.
2531 *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2532 *     to balance a node on its own. These seemed like reasonable ratios.
2533 */
2534static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2535{
2536        unsigned long managed_pages = 0;
2537        unsigned long balanced_pages = 0;
2538        int i;
2539
2540        /* Check the watermark levels */
2541        for (i = 0; i <= classzone_idx; i++) {
2542                struct zone *zone = pgdat->node_zones + i;
2543
2544                if (!populated_zone(zone))
2545                        continue;
2546
2547                managed_pages += zone->managed_pages;
2548
2549                /*
2550                 * A special case here:
2551                 *
2552                 * balance_pgdat() skips over all_unreclaimable after
2553                 * DEF_PRIORITY. Effectively, it considers them balanced so
2554                 * they must be considered balanced here as well!
2555                 */
2556                if (zone->all_unreclaimable) {
2557                        balanced_pages += zone->managed_pages;
2558                        continue;
2559                }
2560
2561                if (zone_balanced(zone, order, 0, i))
2562                        balanced_pages += zone->managed_pages;
2563                else if (!order)
2564                        return false;
2565        }
2566
2567        if (order)
2568                return balanced_pages >= (managed_pages >> 2);
2569        else
2570                return true;
2571}
2572
2573/*
2574 * Prepare kswapd for sleeping. This verifies that there are no processes
2575 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2576 *
2577 * Returns true if kswapd is ready to sleep
2578 */
2579static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2580                                        int classzone_idx)
2581{
2582        /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2583        if (remaining)
2584                return false;
2585
2586        /*
2587         * There is a potential race between when kswapd checks its watermarks
2588         * and a process gets throttled. There is also a potential race if
2589         * processes get throttled, kswapd wakes, a large process exits therby
2590         * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2591         * is going to sleep, no process should be sleeping on pfmemalloc_wait
2592         * so wake them now if necessary. If necessary, processes will wake
2593         * kswapd and get throttled again
2594         */
2595        if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2596                wake_up(&pgdat->pfmemalloc_wait);
2597                return false;
2598        }
2599
2600        return pgdat_balanced(pgdat, order, classzone_idx);
2601}
2602
2603/*
2604 * For kswapd, balance_pgdat() will work across all this node's zones until
2605 * they are all at high_wmark_pages(zone).
2606 *
2607 * Returns the final order kswapd was reclaiming at
2608 *
2609 * There is special handling here for zones which are full of pinned pages.
2610 * This can happen if the pages are all mlocked, or if they are all used by
2611 * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2612 * What we do is to detect the case where all pages in the zone have been
2613 * scanned twice and there has been zero successful reclaim.  Mark the zone as
2614 * dead and from now on, only perform a short scan.  Basically we're polling
2615 * the zone for when the problem goes away.
2616 *
2617 * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2618 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2619 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2620 * lower zones regardless of the number of free pages in the lower zones. This
2621 * interoperates with the page allocator fallback scheme to ensure that aging
2622 * of pages is balanced across the zones.
2623 */
2624static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2625                                                        int *classzone_idx)
2626{
2627        bool pgdat_is_balanced = false;
2628        int i;
2629        int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2630        struct reclaim_state *reclaim_state = current->reclaim_state;
2631        unsigned long nr_soft_reclaimed;
2632        unsigned long nr_soft_scanned;
2633        struct scan_control sc = {
2634                .gfp_mask = GFP_KERNEL,
2635                .may_unmap = 1,
2636                .may_swap = 1,
2637                /*
2638                 * kswapd doesn't want to be bailed out while reclaim. because
2639                 * we want to put equal scanning pressure on each zone.
2640                 */
2641                .nr_to_reclaim = ULONG_MAX,
2642                .order = order,
2643                .target_mem_cgroup = NULL,
2644        };
2645        struct shrink_control shrink = {
2646                .gfp_mask = sc.gfp_mask,
2647        };
2648loop_again:
2649        sc.priority = DEF_PRIORITY;
2650        sc.nr_reclaimed = 0;
2651        sc.may_writepage = !laptop_mode;
2652        count_vm_event(PAGEOUTRUN);
2653
2654        do {
2655                unsigned long lru_pages = 0;
2656
2657                /*
2658                 * Scan in the highmem->dma direction for the highest
2659                 * zone which needs scanning
2660                 */
2661                for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2662                        struct zone *zone = pgdat->node_zones + i;
2663
2664                        if (!populated_zone(zone))
2665                                continue;
2666
2667                        if (zone->all_unreclaimable &&
2668                            sc.priority != DEF_PRIORITY)
2669                                continue;
2670
2671                        /*
2672                         * Do some background aging of the anon list, to give
2673                         * pages a chance to be referenced before reclaiming.
2674                         */
2675                        age_active_anon(zone, &sc);
2676
2677                        /*
2678                         * If the number of buffer_heads in the machine
2679                         * exceeds the maximum allowed level and this node
2680                         * has a highmem zone, force kswapd to reclaim from
2681                         * it to relieve lowmem pressure.
2682                         */
2683                        if (buffer_heads_over_limit && is_highmem_idx(i)) {
2684                                end_zone = i;
2685                                break;
2686                        }
2687
2688                        if (!zone_balanced(zone, order, 0, 0)) {
2689                                end_zone = i;
2690                                break;
2691                        } else {
2692                                /* If balanced, clear the congested flag */
2693                                zone_clear_flag(zone, ZONE_CONGESTED);
2694                        }
2695                }
2696
2697                if (i < 0) {
2698                        pgdat_is_balanced = true;
2699                        goto out;
2700                }
2701
2702                for (i = 0; i <= end_zone; i++) {
2703                        struct zone *zone = pgdat->node_zones + i;
2704
2705                        lru_pages += zone_reclaimable_pages(zone);
2706                }
2707
2708                /*
2709                 * Now scan the zone in the dma->highmem direction, stopping
2710                 * at the last zone which needs scanning.
2711                 *
2712                 * We do this because the page allocator works in the opposite
2713                 * direction.  This prevents the page allocator from allocating
2714                 * pages behind kswapd's direction of progress, which would
2715                 * cause too much scanning of the lower zones.
2716                 */
2717                for (i = 0; i <= end_zone; i++) {
2718                        struct zone *zone = pgdat->node_zones + i;
2719                        int nr_slab, testorder;
2720                        unsigned long balance_gap;
2721
2722                        if (!populated_zone(zone))
2723                                continue;
2724
2725                        if (zone->all_unreclaimable &&
2726                            sc.priority != DEF_PRIORITY)
2727                                continue;
2728
2729                        sc.nr_scanned = 0;
2730
2731                        nr_soft_scanned = 0;
2732                        /*
2733                         * Call soft limit reclaim before calling shrink_zone.
2734                         */
2735                        nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2736                                                        order, sc.gfp_mask,
2737                                                        &nr_soft_scanned);
2738                        sc.nr_reclaimed += nr_soft_reclaimed;
2739
2740                        /*
2741                         * We put equal pressure on every zone, unless
2742                         * one zone has way too many pages free
2743                         * already. The "too many pages" is defined
2744                         * as the high wmark plus a "gap" where the
2745                         * gap is either the low watermark or 1%
2746                         * of the zone, whichever is smaller.
2747                         */
2748                        balance_gap = min(low_wmark_pages(zone),
2749                                (zone->managed_pages +
2750                                        KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2751                                KSWAPD_ZONE_BALANCE_GAP_RATIO);
2752                        /*
2753                         * Kswapd reclaims only single pages with compaction
2754                         * enabled. Trying too hard to reclaim until contiguous
2755                         * free pages have become available can hurt performance
2756                         * by evicting too much useful data from memory.
2757                         * Do not reclaim more than needed for compaction.
2758                         */
2759                        testorder = order;
2760                        if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2761                                        compaction_suitable(zone, order) !=
2762                                                COMPACT_SKIPPED)
2763                                testorder = 0;
2764
2765                        if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
2766                            !zone_balanced(zone, testorder,
2767                                           balance_gap, end_zone)) {
2768                                shrink_zone(zone, &sc);
2769
2770                                reclaim_state->reclaimed_slab = 0;
2771                                nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2772                                sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2773
2774                                if (nr_slab == 0 && !zone_reclaimable(zone))
2775                                        zone->all_unreclaimable = 1;
2776                        }
2777
2778                        /*
2779                         * If we're getting trouble reclaiming, start doing
2780                         * writepage even in laptop mode.
2781                         */
2782                        if (sc.priority < DEF_PRIORITY - 2)
2783                                sc.may_writepage = 1;
2784
2785                        if (zone->all_unreclaimable) {
2786                                if (end_zone && end_zone == i)
2787                                        end_zone--;
2788                                continue;
2789                        }
2790
2791                        if (zone_balanced(zone, testorder, 0, end_zone))
2792                                /*
2793                                 * If a zone reaches its high watermark,
2794                                 * consider it to be no longer congested. It's
2795                                 * possible there are dirty pages backed by
2796                                 * congested BDIs but as pressure is relieved,
2797                                 * speculatively avoid congestion waits
2798                                 */
2799                                zone_clear_flag(zone, ZONE_CONGESTED);
2800                }
2801
2802                /*
2803                 * If the low watermark is met there is no need for processes
2804                 * to be throttled on pfmemalloc_wait as they should not be
2805                 * able to safely make forward progress. Wake them
2806                 */
2807                if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
2808                                pfmemalloc_watermark_ok(pgdat))
2809                        wake_up(&pgdat->pfmemalloc_wait);
2810
2811                if (pgdat_balanced(pgdat, order, *classzone_idx)) {
2812                        pgdat_is_balanced = true;
2813                        break;          /* kswapd: all done */
2814                }
2815
2816                /*
2817                 * We do this so kswapd doesn't build up large priorities for
2818                 * example when it is freeing in parallel with allocators. It
2819                 * matches the direct reclaim path behaviour in terms of impact
2820                 * on zone->*_priority.
2821                 */
2822                if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2823                        break;
2824        } while (--sc.priority >= 0);
2825
2826out:
2827        if (!pgdat_is_balanced) {
2828                cond_resched();
2829
2830                try_to_freeze();
2831
2832                /*
2833                 * Fragmentation may mean that the system cannot be
2834                 * rebalanced for high-order allocations in all zones.
2835                 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2836                 * it means the zones have been fully scanned and are still
2837                 * not balanced. For high-order allocations, there is
2838                 * little point trying all over again as kswapd may
2839                 * infinite loop.
2840                 *
2841                 * Instead, recheck all watermarks at order-0 as they
2842                 * are the most important. If watermarks are ok, kswapd will go
2843                 * back to sleep. High-order users can still perform direct
2844                 * reclaim if they wish.
2845                 */
2846                if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2847                        order = sc.order = 0;
2848
2849                goto loop_again;
2850        }
2851
2852        /*
2853         * If kswapd was reclaiming at a higher order, it has the option of
2854         * sleeping without all zones being balanced. Before it does, it must
2855         * ensure that the watermarks for order-0 on *all* zones are met and
2856         * that the congestion flags are cleared. The congestion flag must
2857         * be cleared as kswapd is the only mechanism that clears the flag
2858         * and it is potentially going to sleep here.
2859         */
2860        if (order) {
2861                int zones_need_compaction = 1;
2862
2863                for (i = 0; i <= end_zone; i++) {
2864                        struct zone *zone = pgdat->node_zones + i;
2865
2866                        if (!populated_zone(zone))
2867                                continue;
2868
2869                        /* Check if the memory needs to be defragmented. */
2870                        if (zone_watermark_ok(zone, order,
2871                                    low_wmark_pages(zone), *classzone_idx, 0))
2872                                zones_need_compaction = 0;
2873                }
2874
2875                if (zones_need_compaction)
2876                        compact_pgdat(pgdat, order);
2877        }
2878
2879        /*
2880         * Return the order we were reclaiming at so prepare_kswapd_sleep()
2881         * makes a decision on the order we were last reclaiming at. However,
2882         * if another caller entered the allocator slow path while kswapd
2883         * was awake, order will remain at the higher level
2884         */
2885        *classzone_idx = end_zone;
2886        return order;
2887}
2888
2889static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2890{
2891        long remaining = 0;
2892        DEFINE_WAIT(wait);
2893
2894        if (freezing(current) || kthread_should_stop())
2895                return;
2896
2897        prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2898
2899        /* Try to sleep for a short interval */
2900        if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
2901                remaining = schedule_timeout(HZ/10);
2902                finish_wait(&pgdat->kswapd_wait, &wait);
2903                prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2904        }
2905
2906        /*
2907         * After a short sleep, check if it was a premature sleep. If not, then
2908         * go fully to sleep until explicitly woken up.
2909         */
2910        if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
2911                trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2912
2913                /*
2914                 * vmstat counters are not perfectly accurate and the estimated
2915                 * value for counters such as NR_FREE_PAGES can deviate from the
2916                 * true value by nr_online_cpus * threshold. To avoid the zone
2917                 * watermarks being breached while under pressure, we reduce the
2918                 * per-cpu vmstat threshold while kswapd is awake and restore
2919                 * them before going back to sleep.
2920                 */
2921                set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2922
2923                /*
2924                 * Compaction records what page blocks it recently failed to
2925                 * isolate pages from and skips them in the future scanning.
2926                 * When kswapd is going to sleep, it is reasonable to assume
2927                 * that pages and compaction may succeed so reset the cache.
2928                 */
2929                reset_isolation_suitable(pgdat);
2930
2931                if (!kthread_should_stop())
2932                        schedule();
2933
2934                set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2935        } else {
2936                if (remaining)
2937                        count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2938                else
2939                        count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2940        }
2941        finish_wait(&pgdat->kswapd_wait, &wait);
2942}
2943
2944/*
2945 * The background pageout daemon, started as a kernel thread
2946 * from the init process.
2947 *
2948 * This basically trickles out pages so that we have _some_
2949 * free memory available even if there is no other activity
2950 * that frees anything up. This is needed for things like routing
2951 * etc, where we otherwise might have all activity going on in
2952 * asynchronous contexts that cannot page things out.
2953 *
2954 * If there are applications that are active memory-allocators
2955 * (most normal use), this basically shouldn't matter.
2956 */
2957static int kswapd(void *p)
2958{
2959        unsigned long order, new_order;
2960        unsigned balanced_order;
2961        int classzone_idx, new_classzone_idx;
2962        int balanced_classzone_idx;
2963        pg_data_t *pgdat = (pg_data_t*)p;
2964        struct task_struct *tsk = current;
2965
2966        struct reclaim_state reclaim_state = {
2967                .reclaimed_slab = 0,
2968        };
2969        const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2970
2971        lockdep_set_current_reclaim_state(GFP_KERNEL);
2972
2973        if (!cpumask_empty(cpumask))
2974                set_cpus_allowed_ptr(tsk, cpumask);
2975        current->reclaim_state = &reclaim_state;
2976
2977        /*
2978         * Tell the memory management that we're a "memory allocator",
2979         * and that if we need more memory we should get access to it
2980         * regardless (see "__alloc_pages()"). "kswapd" should
2981         * never get caught in the normal page freeing logic.
2982         *
2983         * (Kswapd normally doesn't need memory anyway, but sometimes
2984         * you need a small amount of memory in order to be able to
2985         * page out something else, and this flag essentially protects
2986         * us from recursively trying to free more memory as we're
2987         * trying to free the first piece of memory in the first place).
2988         */
2989        tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2990        set_freezable();
2991
2992        order = new_order = 0;
2993        balanced_order = 0;
2994        classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2995        balanced_classzone_idx = classzone_idx;
2996        for ( ; ; ) {
2997                bool ret;
2998
2999                /*
3000                 * If the last balance_pgdat was unsuccessful it's unlikely a
3001                 * new request of a similar or harder type will succeed soon
3002                 * so consider going to sleep on the basis we reclaimed at
3003                 */
3004                if (balanced_classzone_idx >= new_classzone_idx &&
3005                                        balanced_order == new_order) {
3006                        new_order = pgdat->kswapd_max_order;
3007                        new_classzone_idx = pgdat->classzone_idx;
3008                        pgdat->kswapd_max_order =  0;
3009                        pgdat->classzone_idx = pgdat->nr_zones - 1;
3010                }
3011
3012                if (order < new_order || classzone_idx > new_classzone_idx) {
3013                        /*
3014                         * Don't sleep if someone wants a larger 'order'
3015                         * allocation or has tigher zone constraints
3016                         */
3017                        order = new_order;
3018                        classzone_idx = new_classzone_idx;
3019                } else {
3020                        kswapd_try_to_sleep(pgdat, balanced_order,
3021                                                balanced_classzone_idx);
3022                        order = pgdat->kswapd_max_order;
3023                        classzone_idx = pgdat->classzone_idx;
3024                        new_order = order;
3025                        new_classzone_idx = classzone_idx;
3026                        pgdat->kswapd_max_order = 0;
3027                        pgdat->classzone_idx = pgdat->nr_zones - 1;
3028                }
3029
3030                ret = try_to_freeze();
3031                if (kthread_should_stop())
3032                        break;
3033
3034                /*
3035                 * We can speed up thawing tasks if we don't call balance_pgdat
3036                 * after returning from the refrigerator
3037                 */
3038                if (!ret) {
3039                        trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3040                        balanced_classzone_idx = classzone_idx;
3041                        balanced_order = balance_pgdat(pgdat, order,
3042                                                &balanced_classzone_idx);
3043                }
3044        }
3045
3046        current->reclaim_state = NULL;
3047        return 0;
3048}
3049
3050/*
3051 * A zone is low on free memory, so wake its kswapd task to service it.
3052 */
3053void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3054{
3055        pg_data_t *pgdat;
3056
3057        if (!populated_zone(zone))
3058                return;
3059
3060        if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3061                return;
3062        pgdat = zone->zone_pgdat;
3063        if (pgdat->kswapd_max_order < order) {
3064                pgdat->kswapd_max_order = order;
3065                pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3066        }
3067        if (!waitqueue_active(&pgdat->kswapd_wait))
3068                return;
3069        if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3070                return;
3071
3072        trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3073        wake_up_interruptible(&pgdat->kswapd_wait);
3074}
3075
3076/*
3077 * The reclaimable count would be mostly accurate.
3078 * The less reclaimable pages may be
3079 * - mlocked pages, which will be moved to unevictable list when encountered
3080 * - mapped pages, which may require several travels to be reclaimed
3081 * - dirty pages, which is not "instantly" reclaimable
3082 */
3083unsigned long global_reclaimable_pages(void)
3084{
3085        int nr;
3086
3087        nr = global_page_state(NR_ACTIVE_FILE) +
3088             global_page_state(NR_INACTIVE_FILE);
3089
3090        if (get_nr_swap_pages() > 0)
3091                nr += global_page_state(NR_ACTIVE_ANON) +
3092                      global_page_state(NR_INACTIVE_ANON);
3093
3094        return nr;
3095}
3096
3097unsigned long zone_reclaimable_pages(struct zone *zone)
3098{
3099        int nr;
3100
3101        nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3102             zone_page_state(zone, NR_INACTIVE_FILE);
3103
3104        if (get_nr_swap_pages() > 0)
3105                nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3106                      zone_page_state(zone, NR_INACTIVE_ANON);
3107
3108        return nr;
3109}
3110
3111#ifdef CONFIG_HIBERNATION
3112/*
3113 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3114 * freed pages.
3115 *
3116 * Rather than trying to age LRUs the aim is to preserve the overall
3117 * LRU order by reclaiming preferentially
3118 * inactive > active > active referenced > active mapped
3119 */
3120unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3121{
3122        struct reclaim_state reclaim_state;
3123        struct scan_control sc = {
3124                .gfp_mask = GFP_HIGHUSER_MOVABLE,
3125                .may_swap = 1,
3126                .may_unmap = 1,
3127                .may_writepage = 1,
3128                .nr_to_reclaim = nr_to_reclaim,
3129                .hibernation_mode = 1,
3130                .order = 0,
3131                .priority = DEF_PRIORITY,
3132        };
3133        struct shrink_control shrink = {
3134                .gfp_mask = sc.gfp_mask,
3135        };
3136        struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3137        struct task_struct *p = current;
3138        unsigned long nr_reclaimed;
3139
3140        p->flags |= PF_MEMALLOC;
3141        lockdep_set_current_reclaim_state(sc.gfp_mask);
3142        reclaim_state.reclaimed_slab = 0;
3143        p->reclaim_state = &reclaim_state;
3144
3145        nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3146
3147        p->reclaim_state = NULL;
3148        lockdep_clear_current_reclaim_state();
3149        p->flags &= ~PF_MEMALLOC;
3150
3151        return nr_reclaimed;
3152}
3153#endif /* CONFIG_HIBERNATION */
3154
3155/* It's optimal to keep kswapds on the same CPUs as their memory, but
3156   not required for correctness.  So if the last cpu in a node goes
3157   away, we get changed to run anywhere: as the first one comes back,
3158   restore their cpu bindings. */
3159static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3160                        void *hcpu)
3161{
3162        int nid;
3163
3164        if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3165                for_each_node_state(nid, N_MEMORY) {
3166                        pg_data_t *pgdat = NODE_DATA(nid);
3167                        const struct cpumask *mask;
3168
3169                        mask = cpumask_of_node(pgdat->node_id);
3170
3171                        if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3172                                /* One of our CPUs online: restore mask */
3173                                set_cpus_allowed_ptr(pgdat->kswapd, mask);
3174                }
3175        }
3176        return NOTIFY_OK;
3177}
3178
3179/*
3180 * This kswapd start function will be called by init and node-hot-add.
3181 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3182 */
3183int kswapd_run(int nid)
3184{
3185        pg_data_t *pgdat = NODE_DATA(nid);
3186        int ret = 0;
3187
3188        if (pgdat->kswapd)
3189                return 0;
3190
3191        pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3192        if (IS_ERR(pgdat->kswapd)) {
3193                /* failure at boot is fatal */
3194                BUG_ON(system_state == SYSTEM_BOOTING);
3195                pr_err("Failed to start kswapd on node %d\n", nid);
3196                ret = PTR_ERR(pgdat->kswapd);
3197                pgdat->kswapd = NULL;
3198        }
3199        return ret;
3200}
3201
3202/*
3203 * Called by memory hotplug when all memory in a node is offlined.  Caller must
3204 * hold lock_memory_hotplug().
3205 */
3206void kswapd_stop(int nid)
3207{
3208        struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3209
3210        if (kswapd) {
3211                kthread_stop(kswapd);
3212                NODE_DATA(nid)->kswapd = NULL;
3213        }
3214}
3215
3216static int __init kswapd_init(void)
3217{
3218        int nid;
3219
3220        swap_setup();
3221        for_each_node_state(nid, N_MEMORY)
3222                kswapd_run(nid);
3223        hotcpu_notifier(cpu_callback, 0);
3224        return 0;
3225}
3226
3227module_init(kswapd_init)
3228
3229#ifdef CONFIG_NUMA
3230/*
3231 * Zone reclaim mode
3232 *
3233 * If non-zero call zone_reclaim when the number of free pages falls below
3234 * the watermarks.
3235 */
3236int zone_reclaim_mode __read_mostly;
3237
3238#define RECLAIM_OFF 0
3239#define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3240#define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3241#define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3242
3243/*
3244 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3245 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3246 * a zone.
3247 */
3248#define ZONE_RECLAIM_PRIORITY 4
3249
3250/*
3251 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3252 * occur.
3253 */
3254int sysctl_min_unmapped_ratio = 1;
3255
3256/*
3257 * If the number of slab pages in a zone grows beyond this percentage then
3258 * slab reclaim needs to occur.
3259 */
3260int sysctl_min_slab_ratio = 5;
3261
3262static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3263{
3264        unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3265        unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3266                zone_page_state(zone, NR_ACTIVE_FILE);
3267
3268        /*
3269         * It's possible for there to be more file mapped pages than
3270         * accounted for by the pages on the file LRU lists because
3271         * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3272         */
3273        return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3274}
3275
3276/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3277static long zone_pagecache_reclaimable(struct zone *zone)
3278{
3279        long nr_pagecache_reclaimable;
3280        long delta = 0;
3281
3282        /*
3283         * If RECLAIM_SWAP is set, then all file pages are considered
3284         * potentially reclaimable. Otherwise, we have to worry about
3285         * pages like swapcache and zone_unmapped_file_pages() provides
3286         * a better estimate
3287         */
3288        if (zone_reclaim_mode & RECLAIM_SWAP)
3289                nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3290        else
3291                nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3292
3293        /* If we can't clean pages, remove dirty pages from consideration */
3294        if (!(zone_reclaim_mode & RECLAIM_WRITE))
3295                delta += zone_page_state(zone, NR_FILE_DIRTY);
3296
3297        /* Watch for any possible underflows due to delta */
3298        if (unlikely(delta > nr_pagecache_reclaimable))
3299                delta = nr_pagecache_reclaimable;
3300
3301        return nr_pagecache_reclaimable - delta;
3302}
3303
3304/*
3305 * Try to free up some pages from this zone through reclaim.
3306 */
3307static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3308{
3309        /* Minimum pages needed in order to stay on node */
3310        const unsigned long nr_pages = 1 << order;
3311        struct task_struct *p = current;
3312        struct reclaim_state reclaim_state;
3313        struct scan_control sc = {
3314                .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3315                .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3316                .may_swap = 1,
3317                .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3318                .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3319                .order = order,
3320                .priority = ZONE_RECLAIM_PRIORITY,
3321        };
3322        struct shrink_control shrink = {
3323                .gfp_mask = sc.gfp_mask,
3324        };
3325        unsigned long nr_slab_pages0, nr_slab_pages1;
3326
3327        cond_resched();
3328        /*
3329         * We need to be able to allocate from the reserves for RECLAIM_SWAP
3330         * and we also need to be able to write out pages for RECLAIM_WRITE
3331         * and RECLAIM_SWAP.
3332         */
3333        p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3334        lockdep_set_current_reclaim_state(gfp_mask);
3335        reclaim_state.reclaimed_slab = 0;
3336        p->reclaim_state = &reclaim_state;
3337
3338        if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3339                /*
3340                 * Free memory by calling shrink zone with increasing
3341                 * priorities until we have enough memory freed.
3342                 */
3343                do {
3344                        shrink_zone(zone, &sc);
3345                } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3346        }
3347
3348        nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3349        if (nr_slab_pages0 > zone->min_slab_pages) {
3350                /*
3351                 * shrink_slab() does not currently allow us to determine how
3352                 * many pages were freed in this zone. So we take the current
3353                 * number of slab pages and shake the slab until it is reduced
3354                 * by the same nr_pages that we used for reclaiming unmapped
3355                 * pages.
3356                 *
3357                 * Note that shrink_slab will free memory on all zones and may
3358                 * take a long time.
3359                 */
3360                for (;;) {
3361                        unsigned long lru_pages = zone_reclaimable_pages(zone);
3362
3363                        /* No reclaimable slab or very low memory pressure */
3364                        if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3365                                break;
3366
3367                        /* Freed enough memory */
3368                        nr_slab_pages1 = zone_page_state(zone,
3369                                                        NR_SLAB_RECLAIMABLE);
3370                        if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3371                                break;
3372                }
3373
3374                /*
3375                 * Update nr_reclaimed by the number of slab pages we
3376                 * reclaimed from this zone.
3377                 */
3378                nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3379                if (nr_slab_pages1 < nr_slab_pages0)
3380                        sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3381        }
3382
3383        p->reclaim_state = NULL;
3384        current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3385        lockdep_clear_current_reclaim_state();
3386        return sc.nr_reclaimed >= nr_pages;
3387}
3388
3389int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3390{
3391        int node_id;
3392        int ret;
3393
3394        /*
3395         * Zone reclaim reclaims unmapped file backed pages and
3396         * slab pages if we are over the defined limits.
3397         *
3398         * A small portion of unmapped file backed pages is needed for
3399         * file I/O otherwise pages read by file I/O will be immediately
3400         * thrown out if the zone is overallocated. So we do not reclaim
3401         * if less than a specified percentage of the zone is used by
3402         * unmapped file backed pages.
3403         */
3404        if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3405            zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3406                return ZONE_RECLAIM_FULL;
3407
3408        if (zone->all_unreclaimable)
3409                return ZONE_RECLAIM_FULL;
3410
3411        /*
3412         * Do not scan if the allocation should not be delayed.
3413         */
3414        if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3415                return ZONE_RECLAIM_NOSCAN;
3416
3417        /*
3418         * Only run zone reclaim on the local zone or on zones that do not
3419         * have associated processors. This will favor the local processor
3420         * over remote processors and spread off node memory allocations
3421         * as wide as possible.
3422         */
3423        node_id = zone_to_nid(zone);
3424        if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3425                return ZONE_RECLAIM_NOSCAN;
3426
3427        if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3428                return ZONE_RECLAIM_NOSCAN;
3429
3430        ret = __zone_reclaim(zone, gfp_mask, order);
3431        zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3432
3433        if (!ret)
3434                count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3435
3436        return ret;
3437}
3438#endif
3439
3440/*
3441 * page_evictable - test whether a page is evictable
3442 * @page: the page to test
3443 *
3444 * Test whether page is evictable--i.e., should be placed on active/inactive
3445 * lists vs unevictable list.
3446 *
3447 * Reasons page might not be evictable:
3448 * (1) page's mapping marked unevictable
3449 * (2) page is part of an mlocked VMA
3450 *
3451 */
3452int page_evictable(struct page *page)
3453{
3454        return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3455}
3456
3457#ifdef CONFIG_SHMEM
3458/**
3459 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3460 * @pages:      array of pages to check
3461 * @nr_pages:   number of pages to check
3462 *
3463 * Checks pages for evictability and moves them to the appropriate lru list.
3464 *
3465 * This function is only used for SysV IPC SHM_UNLOCK.
3466 */
3467void check_move_unevictable_pages(struct page **pages, int nr_pages)
3468{
3469        struct lruvec *lruvec;
3470        struct zone *zone = NULL;
3471        int pgscanned = 0;
3472        int pgrescued = 0;
3473        int i;
3474
3475        for (i = 0; i < nr_pages; i++) {
3476                struct page *page = pages[i];
3477                struct zone *pagezone;
3478
3479                pgscanned++;
3480                pagezone = page_zone(page);
3481                if (pagezone != zone) {
3482                        if (zone)
3483                                spin_unlock_irq(&zone->lru_lock);
3484                        zone = pagezone;
3485                        spin_lock_irq(&zone->lru_lock);
3486                }
3487                lruvec = mem_cgroup_page_lruvec(page, zone);
3488
3489                if (!PageLRU(page) || !PageUnevictable(page))
3490                        continue;
3491
3492                if (page_evictable(page)) {
3493                        enum lru_list lru = page_lru_base_type(page);
3494
3495                        VM_BUG_ON(PageActive(page));
3496                        ClearPageUnevictable(page);
3497                        del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3498                        add_page_to_lru_list(page, lruvec, lru);
3499                        pgrescued++;
3500                }
3501        }
3502
3503        if (zone) {
3504                __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3505                __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3506                spin_unlock_irq(&zone->lru_lock);
3507        }
3508}
3509#endif /* CONFIG_SHMEM */
3510
3511static void warn_scan_unevictable_pages(void)
3512{
3513        printk_once(KERN_WARNING
3514                    "%s: The scan_unevictable_pages sysctl/node-interface has been "
3515                    "disabled for lack of a legitimate use case.  If you have "
3516                    "one, please send an email to linux-mm@kvack.org.\n",
3517                    current->comm);
3518}
3519
3520/*
3521 * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3522 * all nodes' unevictable lists for evictable pages
3523 */
3524unsigned long scan_unevictable_pages;
3525
3526int scan_unevictable_handler(struct ctl_table *table, int write,
3527                           void __user *buffer,
3528                           size_t *length, loff_t *ppos)
3529{
3530        warn_scan_unevictable_pages();
3531        proc_doulongvec_minmax(table, write, buffer, length, ppos);
3532        scan_unevictable_pages = 0;
3533        return 0;
3534}
3535
3536#ifdef CONFIG_NUMA
3537/*
3538 * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3539 * a specified node's per zone unevictable lists for evictable pages.
3540 */
3541
3542static ssize_t read_scan_unevictable_node(struct device *dev,
3543                                          struct device_attribute *attr,
3544                                          char *buf)
3545{
3546        warn_scan_unevictable_pages();
3547        return sprintf(buf, "0\n");     /* always zero; should fit... */
3548}
3549
3550static ssize_t write_scan_unevictable_node(struct device *dev,
3551                                           struct device_attribute *attr,
3552                                        const char *buf, size_t count)
3553{
3554        warn_scan_unevictable_pages();
3555        return 1;
3556}
3557
3558
3559static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3560                        read_scan_unevictable_node,
3561                        write_scan_unevictable_node);
3562
3563int scan_unevictable_register_node(struct node *node)
3564{
3565        return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3566}
3567
3568void scan_unevictable_unregister_node(struct node *node)
3569{
3570        device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3571}
3572#endif
3573
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.